Skip to content

O
opencv
Commit b2abd8ca authored by Alexander Alekhin's avatar Alexander Alekhin
Browse files
Options

Merge remote-tracking branch 'upstream/3.4' into merge-3.4

parents 2aed9aea b998c06d
Expand all Hide whitespace changes
Inline Side-by-side
Showing with 885 additions and 81 deletions
cmake/OpenCVCompilerDefenses.cmake
View file @ b2abd8ca
......@@ -5,7 +5,7 @@
# - https://wiki.debian.org/Hardening
# - https://wiki.gentoo.org/wiki/Hardened/Toolchain
# - https://docs.microsoft.com/en-us/cpp/build/reference/sdl-enable-additional-security-checks
# - https://developer.apple.com/library/archive/documentation/Security/Conceptual/SecureCodingGuide/Articles/BufferOverflows.html
set(OPENCV_LINKER_DEFENSES_FLAGS_COMMON "")
......@@ -44,6 +44,12 @@ if(MSVC)
if(NOT X86_64)
set(OPENCV_LINKER_DEFENSES_FLAGS_COMMON "${OPENCV_LINKER_DEFENSES_FLAGS_COMMON} /safeseh")
endif()
elseif(CV_CLANG)
ocv_add_defense_compiler_flag("-fstack-protector-strong")
ocv_add_defense_compiler_flag_release("-D_FORTIFY_SOURCE=2")
if (NOT APPLE)
set(OPENCV_LINKER_DEFENSES_FLAGS_COMMON "${OPENCV_LINKER_DEFENSES_FLAGS_COMMON} -z noexecstack -z relro -z now" )
endif()
elseif(CV_GCC)
if(CMAKE_CXX_COMPILER_VERSION VERSION_LESS "4.9")
ocv_add_defense_compiler_flag("-fstack-protector")
......
cmake/OpenCVCompilerOptimizations.cmake
View file @ b2abd8ca
......@@ -4,6 +4,15 @@
# SSE4_1 / SSE4_2 / POPCNT
# AVX / AVX2 / AVX_512F
# FMA3
#
# CPU features groups:
# AVX512_COMMON (Common instructions AVX-512F/CD for all CPUs that support AVX-512)
# AVX512_KNL (Knights Landing with AVX-512F/CD/ER/PF)
# AVX512_KNM (Knights Mill with AVX-512F/CD/ER/PF/4FMAPS/4VNNIW/VPOPCNTDQ)
# AVX512_SKX (Skylake-X with AVX-512F/CD/BW/DQ/VL)
# AVX512_CNL (Cannon Lake with AVX-512F/CD/BW/DQ/VL/IFMA/VBMI)
# AVX512_CEL (Cascade Lake with AVX-512F/CD/BW/DQ/VL/IFMA/VBMI/VNNI)
# AVX512_ICL (Ice Lake with AVX-512F/CD/BW/DQ/VL/IFMA/VBMI/VNNI/VBMI2/BITALG/VPOPCNTDQ/VPCLMULQDQ*/GFNI*/VAES*)
# ppc64le arch:
# VSX (always available on Power8)
......@@ -33,7 +42,8 @@
#
# CPU_{opt}_ENABLED_DEFAULT=ON/OFF - has compiler support without additional flag (CPU_BASELINE_DETECT=ON only)
set(CPU_ALL_OPTIMIZATIONS "SSE;SSE2;SSE3;SSSE3;SSE4_1;SSE4_2;POPCNT;AVX;FP16;AVX2;FMA3;AVX_512F;AVX512_SKX")
set(CPU_ALL_OPTIMIZATIONS "SSE;SSE2;SSE3;SSSE3;SSE4_1;SSE4_2;POPCNT;AVX;FP16;AVX2;FMA3;AVX_512F")
list(APPEND CPU_ALL_OPTIMIZATIONS "AVX512_COMMON;AVX512_KNL;AVX512_KNM;AVX512_SKX;AVX512_CNL;AVX512_CEL;AVX512_ICL")
list(APPEND CPU_ALL_OPTIMIZATIONS NEON VFPV3 FP16)
list(APPEND CPU_ALL_OPTIMIZATIONS VSX VSX3)
list(REMOVE_DUPLICATES CPU_ALL_OPTIMIZATIONS)
......@@ -152,9 +162,15 @@ elseif(" ${CMAKE_CXX_FLAGS} " MATCHES " -march=native | -xHost | /QxHost ")
endif()
if(X86 OR X86_64)
ocv_update(CPU_KNOWN_OPTIMIZATIONS "SSE;SSE2;SSE3;SSSE3;SSE4_1;POPCNT;SSE4_2;FP16;FMA3;AVX;AVX2;AVX_512F;AVX512_SKX")
ocv_update(CPU_KNOWN_OPTIMIZATIONS "SSE;SSE2;SSE3;SSSE3;SSE4_1;POPCNT;SSE4_2;FP16;FMA3;AVX;AVX2;AVX_512F;AVX512_COMMON;AVX512_KNL;AVX512_KNM;AVX512_SKX;AVX512_CNL;AVX512_CEL;AVX512_ICL")
ocv_update(CPU_AVX512_SKX_GROUP "AVX_512F;AVX_512CD;AVX_512BW;AVX_512DQ;AVX_512VL")
ocv_update(CPU_AVX512_COMMON_GROUP "AVX_512F;AVX_512CD")
ocv_update(CPU_AVX512_KNL_GROUP "AVX512_COMMON;AVX512_KNL_EXTRA")
ocv_update(CPU_AVX512_KNM_GROUP "AVX512_KNL;AVX512_KNM_EXTRA;AVX_512VPOPCNTDQ")
ocv_update(CPU_AVX512_SKX_GROUP "AVX512_COMMON;AVX_512VL;AVX_512BW;AVX_512DQ")
ocv_update(CPU_AVX512_CNL_GROUP "AVX512_SKX;AVX_512IFMA;AVX_512VBMI")
ocv_update(CPU_AVX512_CEL_GROUP "AVX512_CNL;AVX_512VNNI")
ocv_update(CPU_AVX512_ICL_GROUP "AVX512_CEL;AVX_512VBMI2;AVX_512BITALG;AVX_512VPOPCNTDQ") # ? VPCLMULQDQ, GFNI, VAES
ocv_update(CPU_SSE_TEST_FILE "${OpenCV_SOURCE_DIR}/cmake/checks/cpu_sse.cpp")
ocv_update(CPU_SSE2_TEST_FILE "${OpenCV_SOURCE_DIR}/cmake/checks/cpu_sse2.cpp")
......@@ -167,9 +183,22 @@ if(X86 OR X86_64)
ocv_update(CPU_AVX2_TEST_FILE "${OpenCV_SOURCE_DIR}/cmake/checks/cpu_avx2.cpp")
ocv_update(CPU_FP16_TEST_FILE "${OpenCV_SOURCE_DIR}/cmake/checks/cpu_fp16.cpp")
ocv_update(CPU_AVX_512F_TEST_FILE "${OpenCV_SOURCE_DIR}/cmake/checks/cpu_avx512.cpp")
ocv_update(CPU_AVX512_COMMON_TEST_FILE "${OpenCV_SOURCE_DIR}/cmake/checks/cpu_avx512common.cpp")
ocv_update(CPU_AVX512_KNL_TEST_FILE "${OpenCV_SOURCE_DIR}/cmake/checks/cpu_avx512knl.cpp")
ocv_update(CPU_AVX512_KNM_TEST_FILE "${OpenCV_SOURCE_DIR}/cmake/checks/cpu_avx512knm.cpp")
ocv_update(CPU_AVX512_SKX_TEST_FILE "${OpenCV_SOURCE_DIR}/cmake/checks/cpu_avx512skx.cpp")
ocv_update(CPU_AVX512_CNL_TEST_FILE "${OpenCV_SOURCE_DIR}/cmake/checks/cpu_avx512cnl.cpp")
ocv_update(CPU_AVX512_CEL_TEST_FILE "${OpenCV_SOURCE_DIR}/cmake/checks/cpu_avx512cel.cpp")
ocv_update(CPU_AVX512_ICL_TEST_FILE "${OpenCV_SOURCE_DIR}/cmake/checks/cpu_avx512icl.cpp")
if(NOT OPENCV_CPU_OPT_IMPLIES_IGNORE)
ocv_update(CPU_AVX512_ICL_IMPLIES "AVX512_CEL")
ocv_update(CPU_AVX512_CEL_IMPLIES "AVX512_CNL")
ocv_update(CPU_AVX512_CNL_IMPLIES "AVX512_SKX")
ocv_update(CPU_AVX512_SKX_IMPLIES "AVX512_COMMON")
ocv_update(CPU_AVX512_KNM_IMPLIES "AVX512_KNL")
ocv_update(CPU_AVX512_KNL_IMPLIES "AVX512_COMMON")
ocv_update(CPU_AVX512_COMMON_IMPLIES "AVX_512F")
ocv_update(CPU_AVX_512F_IMPLIES "AVX2")
ocv_update(CPU_AVX_512F_FORCE "") # Don't force other optimizations
ocv_update(CPU_AVX2_IMPLIES "AVX;FMA3;FP16")
......@@ -192,10 +221,10 @@ if(X86 OR X86_64)
ocv_update(CPU_${name}_FLAGS_NAME "${name}")
if(MSVC)
set(enable_flags "${msvc_flags}")
set(flags_conflict "/arch:[^ ]+")
set(flags_conflict "/arch:[^ ]*|/Qx:[^ ]+")
else()
set(enable_flags "${unix_flags}")
set(flags_conflict "-msse[^ ]*|-mssse3|-mavx[^ ]*|-march[^ ]+")
set(flags_conflict "-msse[^ ]*|-mssse3|-mavx[^ ]*|-march[^ ]*|-x[^ ]+")
endif()
ocv_update(CPU_${name}_FLAGS_ON "${enable_flags}")
if(flags_conflict)
......@@ -215,8 +244,14 @@ if(X86 OR X86_64)
if(NOT X86_64) # x64 compiler doesn't support /arch:sse
ocv_intel_compiler_optimization_option(SSE "-msse" "/arch:SSE")
endif()
ocv_intel_compiler_optimization_option(AVX_512F "-march=common-avx512" "/arch:COMMON-AVX512")
ocv_intel_compiler_optimization_option(AVX512_SKX "-march=core-avx512" "/arch:CORE-AVX512")
ocv_intel_compiler_optimization_option(AVX_512F "-xCOMMON-AVX512" "/Qx:COMMON-AVX512")
ocv_intel_compiler_optimization_option(AVX512_COMMON "-xCOMMON-AVX512" "/Qx:COMMON-AVX512")
ocv_intel_compiler_optimization_option(AVX512_KNL "-xKNL" "/Qx:KNL")
ocv_intel_compiler_optimization_option(AVX512_KNM "-xKNM" "/Qx:KNM")
ocv_intel_compiler_optimization_option(AVX512_SKX "-xSKYLAKE-AVX512" "/Qx:SKYLAKE-AVX512")
ocv_intel_compiler_optimization_option(AVX512_CNL "-xCANNONLAKE" "/Qx:CANNONLAKE")
ocv_intel_compiler_optimization_option(AVX512_CEL "-xCASCADELAKE" "/Qx:CASCADELAKE")
ocv_intel_compiler_optimization_option(AVX512_ICL "-xICELAKE-CLIENT" "/Qx:ICELAKE-CLIENT")
elseif(CV_GCC OR CV_CLANG)
ocv_update(CPU_AVX2_FLAGS_ON "-mavx2")
ocv_update(CPU_FP16_FLAGS_ON "-mf16c")
......@@ -230,12 +265,21 @@ if(X86 OR X86_64)
ocv_update(CPU_SSE2_FLAGS_ON "-msse2")
ocv_update(CPU_SSE_FLAGS_ON "-msse")
if(NOT (CV_GCC AND CMAKE_CXX_COMPILER_VERSION VERSION_LESS "5.0")) # GCC >= 5.0
# -mavx512f -mavx512pf -mavx512er -mavx512cd -mavx512vl -mavx512bw -mavx512dq -mavx512ifma -mavx512vbmi
ocv_update(CPU_AVX_512F_FLAGS_ON "-mavx512f")
ocv_update(CPU_AVX512_SKX_FLAGS_ON "-mavx512f -mavx512cd -mavx512vl -mavx512bw -mavx512dq")
ocv_update(CPU_AVX_512CD_FLAGS_ON "-mavx512cd")
ocv_update(CPU_AVX512_KNL_EXTRA_FLAGS_ON "-mavx512er -mavx512pf")
ocv_update(CPU_AVX512_KNM_EXTRA_FLAGS_ON "-mavx5124fmaps -mavx5124vnniw")
ocv_update(CPU_AVX_512BW_FLAGS_ON "-mavx512bw")
ocv_update(CPU_AVX_512DQ_FLAGS_ON "-mavx512dq")
ocv_update(CPU_AVX_512VL_FLAGS_ON "-mavx512vl")
ocv_update(CPU_AVX_512IFMA_FLAGS_ON "-mavx512ifma")
ocv_update(CPU_AVX_512VBMI_FLAGS_ON "-mavx512vbmi")
ocv_update(CPU_AVX_512VNNI_FLAGS_ON "-mavx512vnni")
ocv_update(CPU_AVX_512VBMI2_FLAGS_ON "-mavx512vbmi2")
ocv_update(CPU_AVX_512BITALG_FLAGS_ON "-mavx512bitalg")
ocv_update(CPU_AVX_512VPOPCNTDQ_FLAGS_ON "-mavx512vpopcntdq")
else()
ocv_update(CPU_AVX_512F_SUPPORTED OFF)
ocv_update(CPU_AVX512_SKX_SUPPORTED OFF)
endif()
elseif(MSVC)
ocv_update(CPU_AVX2_FLAGS_ON "/arch:AVX2")
......@@ -336,6 +380,10 @@ if(CV_DISABLE_OPTIMIZATION)
set(CPU_DISPATCH_REQUIRE "")
endif()
if("x${CPU_DISPATCH}" STREQUAL "xALL")
set(CPU_DISPATCH "${CPU_KNOWN_OPTIMIZATIONS}")
endif()
macro(ocv_check_compiler_optimization OPT)
if(NOT DEFINED CPU_${OPT}_SUPPORTED)
if((DEFINED CPU_${OPT}_FLAGS_ON AND NOT "x${CPU_${OPT}_FLAGS_ON}" STREQUAL "x") OR CPU_${OPT}_TEST_FILE)
......@@ -408,6 +456,7 @@ foreach(OPT ${CPU_KNOWN_OPTIMIZATIONS})
if(NOT DEFINED CPU_${OPT}_FORCE)
set(CPU_${OPT}_FORCE "${CPU_${OPT}_IMPLIES}")
endif()
#message("${OPT}: CPU_${OPT}_FLAGS_ON=${CPU_${OPT}_FLAGS_ON}")
endforeach()
if(_add_native_flag)
......@@ -809,19 +858,19 @@ macro(__ocv_add_dispatched_file filename target_src_var src_directory dst_direct
file(WRITE "${__file}" "${__codestr}")
endif()
if(";${CPU_DISPATCH};" MATCHES "${OPT}" OR __CPU_DISPATCH_INCLUDE_ALL)
if(";${CPU_DISPATCH_FINAL};" MATCHES "${OPT}" OR __CPU_DISPATCH_INCLUDE_ALL)
if(EXISTS "${src_directory}/${filename}.${OPT_LOWER}.cpp")
message(STATUS "Using overrided ${OPT} source: ${src_directory}/${filename}.${OPT_LOWER}.cpp")
else()
list(APPEND ${target_src_var} "${__file}")
endif()
endif()
set(__declarations_str "${__declarations_str}
set(__declarations_str "${__declarations_str}
#define CV_CPU_DISPATCH_MODE ${OPT}
#include \"opencv2/core/private/cv_cpu_include_simd_declarations.hpp\"
")
set(__dispatch_modes "${OPT}, ${__dispatch_modes}")
set(__dispatch_modes "${OPT}, ${__dispatch_modes}")
endif()
endforeach()
set(__declarations_str "${__declarations_str}
......
cmake/checks/cpu_avx512cel.cpp 0 → 100644
View file @ b2abd8ca
#if defined __AVX512__ || defined __AVX512F__
#include <immintrin.h>
void test()
{
__m512i a, b, c;
a = _mm512_dpwssd_epi32(a, b, c);
}
#else
#error "AVX512-CEL is not supported"
#endif
int main() { return 0; }
\ No newline at end of file
cmake/checks/cpu_avx512cnl.cpp 0 → 100644
View file @ b2abd8ca
#if defined __AVX512__ || defined __AVX512F__
#include <immintrin.h>
void test()
{
__m512i a, b, c;
a = _mm512_madd52hi_epu64(a, b, c);
a = _mm512_permutexvar_epi8(a, b);
}
#else
#error "AVX512-CNL is not supported"
#endif
int main() { return 0; }
\ No newline at end of file
cmake/checks/cpu_avx512common.cpp 0 → 100644
View file @ b2abd8ca
#if defined __AVX512__ || defined __AVX512F__
#include <immintrin.h>
void test()
{
__m512i zmm = _mm512_setzero_si512();
zmm = _mm512_lzcnt_epi32(zmm);
#if defined __GNUC__ && defined __x86_64__
asm volatile ("" : : : "zmm16", "zmm17", "zmm18", "zmm19");
#endif
}
#else
#error "AVX512-COMMON is not supported"
#endif
int main() { return 0; }
cmake/checks/cpu_avx512icl.cpp 0 → 100644
View file @ b2abd8ca
#if defined __AVX512__ || defined __AVX512F__
#include <immintrin.h>
void test()
{
__m512i a, b, c;
a = _mm512_popcnt_epi8(a);
a = _mm512_shrdv_epi64(a, b, c);
a = _mm512_popcnt_epi64(a);
}
#else
#error "AVX512-ICL is not supported"
#endif
int main() { return 0; }
\ No newline at end of file
cmake/checks/cpu_avx512knl.cpp 0 → 100644
View file @ b2abd8ca
#if defined __AVX512__ || defined __AVX512F__
#include <immintrin.h>
void test()
{
int* base;
__m512i idx;
__mmask16 m16;
__m512 f;
_mm512_mask_prefetch_i32gather_ps(idx, m16, base, 1, _MM_HINT_T1);
f = _mm512_rsqrt28_ps(f);
}
#else
#error "AVX512-KNL is not supported"
#endif
int main() { return 0; }
\ No newline at end of file
cmake/checks/cpu_avx512knm.cpp 0 → 100644
View file @ b2abd8ca
#if defined __AVX512__ || defined __AVX512F__
#include <immintrin.h>
void test()
{
__m512 a, b, c, d, e;
__m512i ai, bi, ci, di, ei, fi;
__m128 *mem;
__m128i *memi;
__mmask16 m;
a = _mm512_4fnmadd_ps(a, b, c, d, e, mem);
ai = _mm512_4dpwssd_epi32(ai, bi, ci, di, ei, memi);
ai = _mm512_popcnt_epi64(ai);
}
#else
#error "AVX512-KNM is not supported"
#endif
int main() { return 0; }
\ No newline at end of file
doc/opencv.bib
View file @ b2abd8ca
......@@ -195,6 +195,21 @@
volume = {9},
publisher = {Walter de Gruyter}
}
@article{Chaumette06,
author = {Chaumette, Fran{\c c}ois and Hutchinson, S.},
title = {{Visual servo control, Part I: Basic approaches}},
url = {https://hal.inria.fr/inria-00350283},
journal = {{IEEE Robotics and Automation Magazine}},
publisher = {{Institute of Electrical and Electronics Engineers}},
volume = {13},
number = {4},
pages = {82-90},
year = {2006},
pdf = {https://hal.inria.fr/inria-00350283/file/2006_ieee_ram_chaumette.pdf},
hal_id = {inria-00350283},
hal_version = {v1},
}
@article{Daniilidis98,
author = {Konstantinos Daniilidis},
title = {Hand-Eye Calibration Using Dual Quaternions},
......@@ -242,6 +257,12 @@
publisher = {IEEE},
url = {http://alumni.media.mit.edu/~jdavis/Publications/publications_402.pdf}
}
@misc{Eade13,
author = {Eade, Ethan},
title = {Gauss-Newton / Levenberg-Marquardt Optimization},
year = {2013},
url = {http://ethaneade.com/optimization.pdf}
}
@inproceedings{EM11,
author = {Gastal, Eduardo SL and Oliveira, Manuel M},
title = {Domain transform for edge-aware image and video processing},
......@@ -596,10 +617,14 @@
title = {ROF and TV-L1 denoising with Primal-Dual algorithm},
url = {http://znah.net/rof-and-tv-l1-denoising-with-primal-dual-algorithm.html}
}
@misc{VandLec,
author = {Vandenberghe, Lieven},
title = {QR Factorization},
url = {http://www.seas.ucla.edu/~vandenbe/133A/lectures/qr.pdf}
@misc{Madsen04,
author = {K. Madsen and H. B. Nielsen and O. Tingleff},
title = {Methods for Non-Linear Least Squares Problems (2nd ed.)},
year = {2004},
pages = {60},
publisher = {Informatics and Mathematical Modelling, Technical University of Denmark, {DTU}},
address = {Richard Petersens Plads, Building 321, {DK-}2800 Kgs. Lyngby},
url = {http://www2.imm.dtu.dk/pubdb/views/edoc_download.php/3215/pdf/imm3215.pdf}
}
@article{MHT2011,
author = {Getreuer, Pascal},
......@@ -645,6 +670,23 @@
title = {Deeper understanding of the homography decomposition for vision-based control},
year = {2007}
}
@article{Marchand16,
author = {Marchand, Eric and Uchiyama, Hideaki and Spindler, Fabien},
title = {{Pose Estimation for Augmented Reality: A Hands-On Survey}},
url = {https://hal.inria.fr/hal-01246370},
journal = {{IEEE Transactions on Visualization and Computer Graphics}},
publisher = {{Institute of Electrical and Electronics Engineers}},
volume = {22},
number = {12},
pages = {2633 - 2651},
year = {2016},
month = Dec,
doi = {10.1109/TVCG.2015.2513408},
keywords = {homography ; SLAM ; motion estimation ; Index Terms-Survey ; augmented reality ; vision-based camera localization ; pose estimation ; PnP ; keypoint matching ; code examples},
pdf = {https://hal.inria.fr/hal-01246370/file/survey-ieee-v2.pdf},
hal_id = {hal-01246370},
hal_version = {v1},
}
@article{Matas00,
author = {Matas, Jiri and Galambos, Charles and Kittler, Josef},
title = {Robust detection of lines using the progressive probabilistic hough transform},
......@@ -915,6 +957,11 @@
volume = {2},
publisher = {IEEE}
}
@misc{VandLec,
author = {Vandenberghe, Lieven},
title = {QR Factorization},
url = {http://www.seas.ucla.edu/~vandenbe/133A/lectures/qr.pdf}
}
@inproceedings{V03,
author = {Kwatra, Vivek and Sch{\"o}dl, Arno and Essa, Irfan and Turk, Greg and Bobick, Aaron},
title = {Graphcut textures: image and video synthesis using graph cuts},
......
modules/calib3d/include/opencv2/calib3d.hpp
View file @ b2abd8ca
......@@ -383,8 +383,11 @@ public:
modified using setMaxIters() method.
*/
static Ptr<LMSolver> create(const Ptr<LMSolver::Callback>& cb, int maxIters);
static Ptr<LMSolver> create(const Ptr<LMSolver::Callback>& cb, int maxIters, double eps);
};
/** @brief Finds a perspective transformation between two planes.
@param srcPoints Coordinates of the points in the original plane, a matrix of the type CV_32FC2
......@@ -842,6 +845,65 @@ CV_EXPORTS_W int solveP3P( InputArray objectPoints, InputArray imagePoints,
OutputArrayOfArrays rvecs, OutputArrayOfArrays tvecs,
int flags );
/** @brief Refine a pose (the translation and the rotation that transform a 3D point expressed in the object coordinate frame
to the camera coordinate frame) from a 3D-2D point correspondences and starting from an initial solution.
@param objectPoints Array of object points in the object coordinate space, Nx3 1-channel or 1xN/Nx1 3-channel,
where N is the number of points. vector\<Point3f\> can also be passed here.
@param imagePoints Array of corresponding image points, Nx2 1-channel or 1xN/Nx1 2-channel,
where N is the number of points. vector\<Point2f\> can also be passed here.
@param cameraMatrix Input camera matrix \f$A = \vecthreethree{fx}{0}{cx}{0}{fy}{cy}{0}{0}{1}\f$ .
@param distCoeffs Input vector of distortion coefficients
\f$(k_1, k_2, p_1, p_2[, k_3[, k_4, k_5, k_6 [, s_1, s_2, s_3, s_4[, \tau_x, \tau_y]]]])\f$ of
4, 5, 8, 12 or 14 elements. If the vector is NULL/empty, the zero distortion coefficients are
assumed.
@param rvec Input/Output rotation vector (see @ref Rodrigues ) that, together with tvec , brings points from
the model coordinate system to the camera coordinate system. Input values are used as an initial solution.
@param tvec Input/Output translation vector. Input values are used as an initial solution.
@param criteria Criteria when to stop the Levenberg-Marquard iterative algorithm.
The function refines the object pose given at least 3 object points, their corresponding image
projections, an initial solution for the rotation and translation vector,
as well as the camera matrix and the distortion coefficients.
The function minimizes the projection error with respect to the rotation and the translation vectors, according
to a Levenberg-Marquardt iterative minimization @cite Madsen04 @cite Eade13 process.
*/
CV_EXPORTS_W void solvePnPRefineLM( InputArray objectPoints, InputArray imagePoints,
InputArray cameraMatrix, InputArray distCoeffs,
InputOutputArray rvec, InputOutputArray tvec,
TermCriteria criteria = TermCriteria(TermCriteria::EPS + TermCriteria::COUNT, 20, FLT_EPSILON));
/** @brief Refine a pose (the translation and the rotation that transform a 3D point expressed in the object coordinate frame
to the camera coordinate frame) from a 3D-2D point correspondences and starting from an initial solution.
@param objectPoints Array of object points in the object coordinate space, Nx3 1-channel or 1xN/Nx1 3-channel,
where N is the number of points. vector\<Point3f\> can also be passed here.
@param imagePoints Array of corresponding image points, Nx2 1-channel or 1xN/Nx1 2-channel,
where N is the number of points. vector\<Point2f\> can also be passed here.
@param cameraMatrix Input camera matrix \f$A = \vecthreethree{fx}{0}{cx}{0}{fy}{cy}{0}{0}{1}\f$ .
@param distCoeffs Input vector of distortion coefficients
\f$(k_1, k_2, p_1, p_2[, k_3[, k_4, k_5, k_6 [, s_1, s_2, s_3, s_4[, \tau_x, \tau_y]]]])\f$ of
4, 5, 8, 12 or 14 elements. If the vector is NULL/empty, the zero distortion coefficients are
assumed.
@param rvec Input/Output rotation vector (see @ref Rodrigues ) that, together with tvec , brings points from
the model coordinate system to the camera coordinate system. Input values are used as an initial solution.
@param tvec Input/Output translation vector. Input values are used as an initial solution.
@param criteria Criteria when to stop the Levenberg-Marquard iterative algorithm.
@param VVSlambda Gain for the virtual visual servoing control law, equivalent to the \f$\alpha\f$
gain in the Gauss-Newton formulation.
The function refines the object pose given at least 3 object points, their corresponding image
projections, an initial solution for the rotation and translation vector,
as well as the camera matrix and the distortion coefficients.
The function minimizes the projection error with respect to the rotation and the translation vectors, using a
virtual visual servoing (VVS) @cite Chaumette06 @cite Marchand16 scheme.
*/
CV_EXPORTS_W void solvePnPRefineVVS( InputArray objectPoints, InputArray imagePoints,
InputArray cameraMatrix, InputArray distCoeffs,
InputOutputArray rvec, InputOutputArray tvec,
TermCriteria criteria = TermCriteria(TermCriteria::EPS + TermCriteria::COUNT, 20, FLT_EPSILON),
double VVSlambda = 1);
/** @brief Finds an initial camera matrix from 3D-2D point correspondences.
@param objectPoints Vector of vectors of the calibration pattern points in the calibration pattern
......
modules/calib3d/src/levmarq.cpp
View file @ b2abd8ca
......@@ -81,11 +81,11 @@ class LMSolverImpl CV_FINAL : public LMSolver
{
public:
LMSolverImpl() : maxIters(100) { init(); }
LMSolverImpl(const Ptr<LMSolver::Callback>& _cb, int _maxIters) : cb(_cb), maxIters(_maxIters) { init(); }
LMSolverImpl(const Ptr<LMSolver::Callback>& _cb, int _maxIters) : cb(_cb), epsx(FLT_EPSILON), epsf(FLT_EPSILON), maxIters(_maxIters) { init(); }
LMSolverImpl(const Ptr<LMSolver::Callback>& _cb, int _maxIters, double _eps) : cb(_cb), epsx(_eps), epsf(_eps), maxIters(_maxIters) { init(); }
void init()
{
epsx = epsf = FLT_EPSILON;
printInterval = 0;
}
......@@ -215,4 +215,9 @@ Ptr<LMSolver> LMSolver::create(const Ptr<LMSolver::Callback>& cb, int maxIters)
return makePtr<LMSolverImpl>(cb, maxIters);
}
Ptr<LMSolver> LMSolver::create(const Ptr<LMSolver::Callback>& cb, int maxIters, double eps)
{
return makePtr<LMSolverImpl>(cb, maxIters, eps);
}
}
modules/calib3d/src/solvepnp.cpp
View file @ b2abd8ca
......@@ -456,4 +456,271 @@ int solveP3P( InputArray _opoints, InputArray _ipoints,
return solutions;
}
class SolvePnPRefineLMCallback CV_FINAL : public LMSolver::Callback
{
public:
SolvePnPRefineLMCallback(InputArray _opoints, InputArray _ipoints, InputArray _cameraMatrix, InputArray _distCoeffs)
{
objectPoints = _opoints.getMat();
imagePoints = _ipoints.getMat();
npoints = std::max(objectPoints.checkVector(3, CV_32F), objectPoints.checkVector(3, CV_64F));
imagePoints0 = imagePoints.reshape(1, npoints*2);
cameraMatrix = _cameraMatrix.getMat();
distCoeffs = _distCoeffs.getMat();
}
bool compute(InputArray _param, OutputArray _err, OutputArray _Jac) const CV_OVERRIDE
{
Mat param = _param.getMat();
_err.create(npoints*2, 1, CV_64FC1);
if(_Jac.needed())
{
_Jac.create(npoints*2, param.rows, CV_64FC1);
}
Mat rvec = param(Rect(0, 0, 1, 3)), tvec = param(Rect(0, 3, 1, 3));
Mat J, projectedPts;
projectPoints(objectPoints, rvec, tvec, cameraMatrix, distCoeffs, projectedPts, _Jac.needed() ? J : noArray());
if (_Jac.needed())
{
Mat Jac = _Jac.getMat();
for (int i = 0; i < Jac.rows; i++)
{
for (int j = 0; j < Jac.cols; j++)
{
Jac.at<double>(i,j) = J.at<double>(i,j);
}
}
}
Mat err = _err.getMat();
projectedPts = projectedPts.reshape(1, npoints*2);
err = projectedPts - imagePoints0;
return true;
}
Mat objectPoints, imagePoints, imagePoints0;
Mat cameraMatrix, distCoeffs;
int npoints;
};
/**
* @brief Compute the Interaction matrix and the residuals for the current pose.
* @param objectPoints 3D object points.
* @param R Current estimated rotation matrix.
* @param tvec Current estimated translation vector.
* @param L Interaction matrix for a vector of point features.
* @param s Residuals.
*/
static void computeInteractionMatrixAndResiduals(const Mat& objectPoints, const Mat& R, const Mat& tvec,
Mat& L, Mat& s)
{
Mat objectPointsInCam;
int npoints = objectPoints.rows;
for (int i = 0; i < npoints; i++)
{
Mat curPt = objectPoints.row(i);
objectPointsInCam = R * curPt.t() + tvec;
double Zi = objectPointsInCam.at<double>(2,0);
double xi = objectPointsInCam.at<double>(0,0) / Zi;
double yi = objectPointsInCam.at<double>(1,0) / Zi;
s.at<double>(2*i,0) = xi;
s.at<double>(2*i+1,0) = yi;
L.at<double>(2*i,0) = -1 / Zi;
L.at<double>(2*i,1) = 0;
L.at<double>(2*i,2) = xi / Zi;
L.at<double>(2*i,3) = xi*yi;
L.at<double>(2*i,4) = -(1 + xi*xi);
L.at<double>(2*i,5) = yi;
L.at<double>(2*i+1,0) = 0;
L.at<double>(2*i+1,1) = -1 / Zi;
L.at<double>(2*i+1,2) = yi / Zi;
L.at<double>(2*i+1,3) = 1 + yi*yi;
L.at<double>(2*i+1,4) = -xi*yi;
L.at<double>(2*i+1,5) = -xi;
}
}
/**
* @brief The exponential map from se(3) to SE(3).
* @param twist A twist (v, w) represents the velocity of a rigid body as an angular velocity
* around an axis and a linear velocity along this axis.
* @param R1 Resultant rotation matrix from the twist.
* @param t1 Resultant translation vector from the twist.
*/
static void exponentialMapToSE3Inv(const Mat& twist, Mat& R1, Mat& t1)
{
//see Exponential Map in http://ethaneade.com/lie.pdf
/*
\begin{align*}
\boldsymbol{\delta} &= \left( \mathbf{u}, \boldsymbol{\omega} \right ) \in se(3) \\
\mathbf{u}, \boldsymbol{\omega} &\in \mathbb{R}^3 \\
\theta &= \sqrt{ \boldsymbol{\omega}^T \boldsymbol{\omega} } \\
A &= \frac{\sin \theta}{\theta} \\
B &= \frac{1 - \cos \theta}{\theta^2} \\
C &= \frac{1-A}{\theta^2} \\
\mathbf{R} &= \mathbf{I} + A \boldsymbol{\omega}_{\times} + B \boldsymbol{\omega}_{\times}^2 \\
\mathbf{V} &= \mathbf{I} + B \boldsymbol{\omega}_{\times} + C \boldsymbol{\omega}_{\times}^2 \\
\exp \begin{pmatrix}
\mathbf{u} \\
\boldsymbol{\omega}
\end{pmatrix} &=
\left(
\begin{array}{c|c}
\mathbf{R} & \mathbf{V} \mathbf{u} \\ \hline
\mathbf{0} & 1
\end{array}
\right )
\end{align*}
*/
double vx = twist.at<double>(0,0);
double vy = twist.at<double>(1,0);
double vz = twist.at<double>(2,0);
double wx = twist.at<double>(3,0);
double wy = twist.at<double>(4,0);
double wz = twist.at<double>(5,0);
Matx31d rvec(wx, wy, wz);
Mat R;
Rodrigues(rvec, R);
double theta = sqrt(wx*wx + wy*wy + wz*wz);
double sinc = std::fabs(theta) < 1e-8 ? 1 : sin(theta) / theta;
double mcosc = (std::fabs(theta) < 1e-8) ? 0.5 : (1-cos(theta)) / (theta*theta);
double msinc = (std::abs(theta) < 1e-8) ? (1/6.0) : (1-sinc) / (theta*theta);
Matx31d dt;
dt(0) = vx*(sinc + wx*wx*msinc) + vy*(wx*wy*msinc - wz*mcosc) + vz*(wx*wz*msinc + wy*mcosc);
dt(1) = vx*(wx*wy*msinc + wz*mcosc) + vy*(sinc + wy*wy*msinc) + vz*(wy*wz*msinc - wx*mcosc);
dt(2) = vx*(wx*wz*msinc - wy*mcosc) + vy*(wy*wz*msinc + wx*mcosc) + vz*(sinc + wz*wz*msinc);
R1 = R.t();
t1 = -R1 * dt;
}
enum SolvePnPRefineMethod {
SOLVEPNP_REFINE_LM = 0,
SOLVEPNP_REFINE_VVS = 1
};
static void solvePnPRefine(InputArray _objectPoints, InputArray _imagePoints,
InputArray _cameraMatrix, InputArray _distCoeffs,
InputOutputArray _rvec, InputOutputArray _tvec,
SolvePnPRefineMethod _flags,
TermCriteria _criteria=TermCriteria(TermCriteria::EPS+TermCriteria::COUNT, 20, FLT_EPSILON),
double _vvslambda=1)
{
CV_INSTRUMENT_REGION();
Mat opoints_ = _objectPoints.getMat(), ipoints_ = _imagePoints.getMat();
Mat opoints, ipoints;
opoints_.convertTo(opoints, CV_64F);
ipoints_.convertTo(ipoints, CV_64F);
int npoints = opoints.checkVector(3, CV_64F);
CV_Assert( npoints >= 3 && npoints == ipoints.checkVector(2, CV_64F) );
CV_Assert( !_rvec.empty() && !_tvec.empty() );
int rtype = _rvec.type(), ttype = _tvec.type();
Size rsize = _rvec.size(), tsize = _tvec.size();
CV_Assert( (rtype == CV_32FC1 || rtype == CV_64FC1) &&
(ttype == CV_32FC1 || ttype == CV_64FC1) );
CV_Assert( (rsize == Size(1, 3) || rsize == Size(3, 1)) &&
(tsize == Size(1, 3) || tsize == Size(3, 1)) );
Mat cameraMatrix0 = _cameraMatrix.getMat();
Mat distCoeffs0 = _distCoeffs.getMat();
Mat cameraMatrix = Mat_<double>(cameraMatrix0);
Mat distCoeffs = Mat_<double>(distCoeffs0);
if (_flags == SOLVEPNP_REFINE_LM)
{
Mat rvec0 = _rvec.getMat(), tvec0 = _tvec.getMat();
Mat rvec, tvec;
rvec0.convertTo(rvec, CV_64F);
tvec0.convertTo(tvec, CV_64F);
Mat params(6, 1, CV_64FC1);
for (int i = 0; i < 3; i++)
{
params.at<double>(i,0) = rvec.at<double>(i,0);
params.at<double>(i+3,0) = tvec.at<double>(i,0);
}
LMSolver::create(makePtr<SolvePnPRefineLMCallback>(opoints, ipoints, cameraMatrix, distCoeffs), _criteria.maxCount, _criteria.epsilon)->run(params);
params.rowRange(0, 3).convertTo(rvec0, rvec0.depth());
params.rowRange(3, 6).convertTo(tvec0, tvec0.depth());
}
else if (_flags == SOLVEPNP_REFINE_VVS)
{
Mat rvec0 = _rvec.getMat(), tvec0 = _tvec.getMat();
Mat rvec, tvec;
rvec0.convertTo(rvec, CV_64F);
tvec0.convertTo(tvec, CV_64F);
vector<Point2d> ipoints_normalized;
undistortPoints(ipoints, ipoints_normalized, cameraMatrix, distCoeffs);
Mat sd = Mat(ipoints_normalized).reshape(1, npoints*2);
Mat objectPoints0 = opoints.reshape(1, npoints);
Mat imagePoints0 = ipoints.reshape(1, npoints*2);
Mat L(npoints*2, 6, CV_64FC1), s(npoints*2, 1, CV_64FC1);
double residuals_1 = std::numeric_limits<double>::max(), residuals = 0;
Mat err;
Mat R;
Rodrigues(rvec, R);
for (int iter = 0; iter < _criteria.maxCount; iter++)
{
computeInteractionMatrixAndResiduals(objectPoints0, R, tvec, L, s);
err = s - sd;
Mat Lp = L.inv(cv::DECOMP_SVD);
Mat dq = -_vvslambda * Lp * err;
Mat R1, t1;
exponentialMapToSE3Inv(dq, R1, t1);
R = R1 * R;
tvec = R1 * tvec + t1;
residuals_1 = residuals;
Mat res = err.t()*err;
residuals = res.at<double>(0,0);
if (std::fabs(residuals - residuals_1) < _criteria.epsilon)
break;
}
Rodrigues(R, rvec);
rvec.convertTo(rvec0, rvec0.depth());
tvec.convertTo(tvec0, tvec0.depth());
}
}
void solvePnPRefineLM(InputArray _objectPoints, InputArray _imagePoints,
InputArray _cameraMatrix, InputArray _distCoeffs,
InputOutputArray _rvec, InputOutputArray _tvec,
TermCriteria _criteria)
{
CV_INSTRUMENT_REGION();
solvePnPRefine(_objectPoints, _imagePoints, _cameraMatrix, _distCoeffs, _rvec, _tvec, SOLVEPNP_REFINE_LM, _criteria);
}
void solvePnPRefineVVS(InputArray _objectPoints, InputArray _imagePoints,
InputArray _cameraMatrix, InputArray _distCoeffs,
InputOutputArray _rvec, InputOutputArray _tvec,
TermCriteria _criteria, double _VVSlambda)
{
CV_INSTRUMENT_REGION();
solvePnPRefine(_objectPoints, _imagePoints, _cameraMatrix, _distCoeffs, _rvec, _tvec, SOLVEPNP_REFINE_VVS, _criteria, _VVSlambda);
}
}
modules/core/include/opencv2/core/cv_cpu_dispatch.h
View file @ b2abd8ca
......@@ -87,9 +87,41 @@
# include <immintrin.h>
# define CV_AVX_512F 1
#endif
#ifdef CV_CPU_COMPILE_AVX512_COMMON
# define CV_AVX512_COMMON 1
# define CV_AVX_512CD 1
#endif
#ifdef CV_CPU_COMPILE_AVX512_KNL
# define CV_AVX512_KNL 1
# define CV_AVX_512ER 1
# define CV_AVX_512PF 1
#endif
#ifdef CV_CPU_COMPILE_AVX512_KNM
# define CV_AVX512_KNM 1
# define CV_AVX_5124FMAPS 1
# define CV_AVX_5124VNNIW 1
# define CV_AVX_512VPOPCNTDQ 1
#endif
#ifdef CV_CPU_COMPILE_AVX512_SKX
# include <immintrin.h>
# define CV_AVX512_SKX 1
# define CV_AVX_512VL 1
# define CV_AVX_512BW 1
# define CV_AVX_512DQ 1
#endif
#ifdef CV_CPU_COMPILE_AVX512_CNL
# define CV_AVX512_CNL 1
# define CV_AVX_512IFMA 1
# define CV_AVX_512VBMI 1
#endif
#ifdef CV_CPU_COMPILE_AVX512_CEL
# define CV_AVX512_CEL 1
# define CV_AVX_512VNNI 1
#endif
#ifdef CV_CPU_COMPILE_AVX512_ICL
# define CV_AVX512_ICL 1
# define CV_AVX_512VBMI2 1
# define CV_AVX_512BITALG 1
# define CV_AVX_512VPOPCNTDQ 1
#endif
#ifdef CV_CPU_COMPILE_FMA3
# define CV_FMA3 1
......@@ -223,9 +255,10 @@ struct VZeroUpperGuard {
#ifndef CV_AVX_512ER
# define CV_AVX_512ER 0
#endif
#ifndef CV_AVX_512IFMA512
# define CV_AVX_512IFMA512 0
#ifndef CV_AVX_512IFMA
# define CV_AVX_512IFMA 0
#endif
#define CV_AVX_512IFMA512 CV_AVX_512IFMA // deprecated
#ifndef CV_AVX_512PF
# define CV_AVX_512PF 0
#endif
......@@ -235,9 +268,45 @@ struct VZeroUpperGuard {
#ifndef CV_AVX_512VL
# define CV_AVX_512VL 0
#endif
#ifndef CV_AVX_5124FMAPS
# define CV_AVX_5124FMAPS 0
#endif
#ifndef CV_AVX_5124VNNIW
# define CV_AVX_5124VNNIW 0
#endif
#ifndef CV_AVX_512VPOPCNTDQ
# define CV_AVX_512VPOPCNTDQ 0
#endif
#ifndef CV_AVX_512VNNI
# define CV_AVX_512VNNI 0
#endif
#ifndef CV_AVX_512VBMI2
# define CV_AVX_512VBMI2 0
#endif
#ifndef CV_AVX_512BITALG
# define CV_AVX_512BITALG 0
#endif
#ifndef CV_AVX512_COMMON
# define CV_AVX512_COMMON 0
#endif
#ifndef CV_AVX512_KNL
# define CV_AVX512_KNL 0
#endif
#ifndef CV_AVX512_KNM
# define CV_AVX512_KNM 0
#endif
#ifndef CV_AVX512_SKX
# define CV_AVX512_SKX 0
#endif
#ifndef CV_AVX512_CNL
# define CV_AVX512_CNL 0
#endif
#ifndef CV_AVX512_CEL
# define CV_AVX512_CEL 0
#endif
#ifndef CV_AVX512_ICL
# define CV_AVX512_ICL 0
#endif
#ifndef CV_NEON
# define CV_NEON 0
......
modules/core/include/opencv2/core/cv_cpu_helper.h
View file @ b2abd8ca
......@@ -252,6 +252,69 @@
#endif
#define __CV_CPU_DISPATCH_CHAIN_AVX_512F(fn, args, mode, ...) CV_CPU_CALL_AVX_512F(fn, args); __CV_EXPAND(__CV_CPU_DISPATCH_CHAIN_ ## mode(fn, args, __VA_ARGS__))
#if !defined CV_DISABLE_OPTIMIZATION && defined CV_ENABLE_INTRINSICS && defined CV_CPU_COMPILE_AVX512_COMMON
# define CV_TRY_AVX512_COMMON 1
# define CV_CPU_FORCE_AVX512_COMMON 1
# define CV_CPU_HAS_SUPPORT_AVX512_COMMON 1
# define CV_CPU_CALL_AVX512_COMMON(fn, args) return (cpu_baseline::fn args)
# define CV_CPU_CALL_AVX512_COMMON_(fn, args) return (opt_AVX512_COMMON::fn args)
#elif !defined CV_DISABLE_OPTIMIZATION && defined CV_ENABLE_INTRINSICS && defined CV_CPU_DISPATCH_COMPILE_AVX512_COMMON
# define CV_TRY_AVX512_COMMON 1
# define CV_CPU_FORCE_AVX512_COMMON 0
# define CV_CPU_HAS_SUPPORT_AVX512_COMMON (cv::checkHardwareSupport(CV_CPU_AVX512_COMMON))
# define CV_CPU_CALL_AVX512_COMMON(fn, args) if (CV_CPU_HAS_SUPPORT_AVX512_COMMON) return (opt_AVX512_COMMON::fn args)
# define CV_CPU_CALL_AVX512_COMMON_(fn, args) if (CV_CPU_HAS_SUPPORT_AVX512_COMMON) return (opt_AVX512_COMMON::fn args)
#else
# define CV_TRY_AVX512_COMMON 0
# define CV_CPU_FORCE_AVX512_COMMON 0
# define CV_CPU_HAS_SUPPORT_AVX512_COMMON 0
# define CV_CPU_CALL_AVX512_COMMON(fn, args)
# define CV_CPU_CALL_AVX512_COMMON_(fn, args)
#endif
#define __CV_CPU_DISPATCH_CHAIN_AVX512_COMMON(fn, args, mode, ...) CV_CPU_CALL_AVX512_COMMON(fn, args); __CV_EXPAND(__CV_CPU_DISPATCH_CHAIN_ ## mode(fn, args, __VA_ARGS__))
#if !defined CV_DISABLE_OPTIMIZATION && defined CV_ENABLE_INTRINSICS && defined CV_CPU_COMPILE_AVX512_KNL
# define CV_TRY_AVX512_KNL 1
# define CV_CPU_FORCE_AVX512_KNL 1
# define CV_CPU_HAS_SUPPORT_AVX512_KNL 1
# define CV_CPU_CALL_AVX512_KNL(fn, args) return (cpu_baseline::fn args)
# define CV_CPU_CALL_AVX512_KNL_(fn, args) return (opt_AVX512_KNL::fn args)
#elif !defined CV_DISABLE_OPTIMIZATION && defined CV_ENABLE_INTRINSICS && defined CV_CPU_DISPATCH_COMPILE_AVX512_KNL
# define CV_TRY_AVX512_KNL 1
# define CV_CPU_FORCE_AVX512_KNL 0
# define CV_CPU_HAS_SUPPORT_AVX512_KNL (cv::checkHardwareSupport(CV_CPU_AVX512_KNL))
# define CV_CPU_CALL_AVX512_KNL(fn, args) if (CV_CPU_HAS_SUPPORT_AVX512_KNL) return (opt_AVX512_KNL::fn args)
# define CV_CPU_CALL_AVX512_KNL_(fn, args) if (CV_CPU_HAS_SUPPORT_AVX512_KNL) return (opt_AVX512_KNL::fn args)
#else
# define CV_TRY_AVX512_KNL 0
# define CV_CPU_FORCE_AVX512_KNL 0
# define CV_CPU_HAS_SUPPORT_AVX512_KNL 0
# define CV_CPU_CALL_AVX512_KNL(fn, args)
# define CV_CPU_CALL_AVX512_KNL_(fn, args)
#endif
#define __CV_CPU_DISPATCH_CHAIN_AVX512_KNL(fn, args, mode, ...) CV_CPU_CALL_AVX512_KNL(fn, args); __CV_EXPAND(__CV_CPU_DISPATCH_CHAIN_ ## mode(fn, args, __VA_ARGS__))
#if !defined CV_DISABLE_OPTIMIZATION && defined CV_ENABLE_INTRINSICS && defined CV_CPU_COMPILE_AVX512_KNM
# define CV_TRY_AVX512_KNM 1
# define CV_CPU_FORCE_AVX512_KNM 1
# define CV_CPU_HAS_SUPPORT_AVX512_KNM 1
# define CV_CPU_CALL_AVX512_KNM(fn, args) return (cpu_baseline::fn args)
# define CV_CPU_CALL_AVX512_KNM_(fn, args) return (opt_AVX512_KNM::fn args)
#elif !defined CV_DISABLE_OPTIMIZATION && defined CV_ENABLE_INTRINSICS && defined CV_CPU_DISPATCH_COMPILE_AVX512_KNM
# define CV_TRY_AVX512_KNM 1
# define CV_CPU_FORCE_AVX512_KNM 0
# define CV_CPU_HAS_SUPPORT_AVX512_KNM (cv::checkHardwareSupport(CV_CPU_AVX512_KNM))
# define CV_CPU_CALL_AVX512_KNM(fn, args) if (CV_CPU_HAS_SUPPORT_AVX512_KNM) return (opt_AVX512_KNM::fn args)
# define CV_CPU_CALL_AVX512_KNM_(fn, args) if (CV_CPU_HAS_SUPPORT_AVX512_KNM) return (opt_AVX512_KNM::fn args)
#else
# define CV_TRY_AVX512_KNM 0
# define CV_CPU_FORCE_AVX512_KNM 0
# define CV_CPU_HAS_SUPPORT_AVX512_KNM 0
# define CV_CPU_CALL_AVX512_KNM(fn, args)
# define CV_CPU_CALL_AVX512_KNM_(fn, args)
#endif
#define __CV_CPU_DISPATCH_CHAIN_AVX512_KNM(fn, args, mode, ...) CV_CPU_CALL_AVX512_KNM(fn, args); __CV_EXPAND(__CV_CPU_DISPATCH_CHAIN_ ## mode(fn, args, __VA_ARGS__))
#if !defined CV_DISABLE_OPTIMIZATION && defined CV_ENABLE_INTRINSICS && defined CV_CPU_COMPILE_AVX512_SKX
# define CV_TRY_AVX512_SKX 1
# define CV_CPU_FORCE_AVX512_SKX 1
......@@ -273,6 +336,69 @@
#endif
#define __CV_CPU_DISPATCH_CHAIN_AVX512_SKX(fn, args, mode, ...) CV_CPU_CALL_AVX512_SKX(fn, args); __CV_EXPAND(__CV_CPU_DISPATCH_CHAIN_ ## mode(fn, args, __VA_ARGS__))
#if !defined CV_DISABLE_OPTIMIZATION && defined CV_ENABLE_INTRINSICS && defined CV_CPU_COMPILE_AVX512_CNL
# define CV_TRY_AVX512_CNL 1
# define CV_CPU_FORCE_AVX512_CNL 1
# define CV_CPU_HAS_SUPPORT_AVX512_CNL 1
# define CV_CPU_CALL_AVX512_CNL(fn, args) return (cpu_baseline::fn args)
# define CV_CPU_CALL_AVX512_CNL_(fn, args) return (opt_AVX512_CNL::fn args)
#elif !defined CV_DISABLE_OPTIMIZATION && defined CV_ENABLE_INTRINSICS && defined CV_CPU_DISPATCH_COMPILE_AVX512_CNL
# define CV_TRY_AVX512_CNL 1
# define CV_CPU_FORCE_AVX512_CNL 0
# define CV_CPU_HAS_SUPPORT_AVX512_CNL (cv::checkHardwareSupport(CV_CPU_AVX512_CNL))
# define CV_CPU_CALL_AVX512_CNL(fn, args) if (CV_CPU_HAS_SUPPORT_AVX512_CNL) return (opt_AVX512_CNL::fn args)
# define CV_CPU_CALL_AVX512_CNL_(fn, args) if (CV_CPU_HAS_SUPPORT_AVX512_CNL) return (opt_AVX512_CNL::fn args)
#else
# define CV_TRY_AVX512_CNL 0
# define CV_CPU_FORCE_AVX512_CNL 0
# define CV_CPU_HAS_SUPPORT_AVX512_CNL 0
# define CV_CPU_CALL_AVX512_CNL(fn, args)
# define CV_CPU_CALL_AVX512_CNL_(fn, args)
#endif
#define __CV_CPU_DISPATCH_CHAIN_AVX512_CNL(fn, args, mode, ...) CV_CPU_CALL_AVX512_CNL(fn, args); __CV_EXPAND(__CV_CPU_DISPATCH_CHAIN_ ## mode(fn, args, __VA_ARGS__))
#if !defined CV_DISABLE_OPTIMIZATION && defined CV_ENABLE_INTRINSICS && defined CV_CPU_COMPILE_AVX512_CEL
# define CV_TRY_AVX512_CEL 1
# define CV_CPU_FORCE_AVX512_CEL 1
# define CV_CPU_HAS_SUPPORT_AVX512_CEL 1
# define CV_CPU_CALL_AVX512_CEL(fn, args) return (cpu_baseline::fn args)
# define CV_CPU_CALL_AVX512_CEL_(fn, args) return (opt_AVX512_CEL::fn args)
#elif !defined CV_DISABLE_OPTIMIZATION && defined CV_ENABLE_INTRINSICS && defined CV_CPU_DISPATCH_COMPILE_AVX512_CEL
# define CV_TRY_AVX512_CEL 1
# define CV_CPU_FORCE_AVX512_CEL 0
# define CV_CPU_HAS_SUPPORT_AVX512_CEL (cv::checkHardwareSupport(CV_CPU_AVX512_CEL))
# define CV_CPU_CALL_AVX512_CEL(fn, args) if (CV_CPU_HAS_SUPPORT_AVX512_CEL) return (opt_AVX512_CEL::fn args)
# define CV_CPU_CALL_AVX512_CEL_(fn, args) if (CV_CPU_HAS_SUPPORT_AVX512_CEL) return (opt_AVX512_CEL::fn args)
#else
# define CV_TRY_AVX512_CEL 0
# define CV_CPU_FORCE_AVX512_CEL 0
# define CV_CPU_HAS_SUPPORT_AVX512_CEL 0
# define CV_CPU_CALL_AVX512_CEL(fn, args)
# define CV_CPU_CALL_AVX512_CEL_(fn, args)
#endif
#define __CV_CPU_DISPATCH_CHAIN_AVX512_CEL(fn, args, mode, ...) CV_CPU_CALL_AVX512_CEL(fn, args); __CV_EXPAND(__CV_CPU_DISPATCH_CHAIN_ ## mode(fn, args, __VA_ARGS__))
#if !defined CV_DISABLE_OPTIMIZATION && defined CV_ENABLE_INTRINSICS && defined CV_CPU_COMPILE_AVX512_ICL
# define CV_TRY_AVX512_ICL 1
# define CV_CPU_FORCE_AVX512_ICL 1
# define CV_CPU_HAS_SUPPORT_AVX512_ICL 1
# define CV_CPU_CALL_AVX512_ICL(fn, args) return (cpu_baseline::fn args)
# define CV_CPU_CALL_AVX512_ICL_(fn, args) return (opt_AVX512_ICL::fn args)
#elif !defined CV_DISABLE_OPTIMIZATION && defined CV_ENABLE_INTRINSICS && defined CV_CPU_DISPATCH_COMPILE_AVX512_ICL
# define CV_TRY_AVX512_ICL 1
# define CV_CPU_FORCE_AVX512_ICL 0
# define CV_CPU_HAS_SUPPORT_AVX512_ICL (cv::checkHardwareSupport(CV_CPU_AVX512_ICL))
# define CV_CPU_CALL_AVX512_ICL(fn, args) if (CV_CPU_HAS_SUPPORT_AVX512_ICL) return (opt_AVX512_ICL::fn args)
# define CV_CPU_CALL_AVX512_ICL_(fn, args) if (CV_CPU_HAS_SUPPORT_AVX512_ICL) return (opt_AVX512_ICL::fn args)
#else
# define CV_TRY_AVX512_ICL 0
# define CV_CPU_FORCE_AVX512_ICL 0
# define CV_CPU_HAS_SUPPORT_AVX512_ICL 0
# define CV_CPU_CALL_AVX512_ICL(fn, args)
# define CV_CPU_CALL_AVX512_ICL_(fn, args)
#endif
#define __CV_CPU_DISPATCH_CHAIN_AVX512_ICL(fn, args, mode, ...) CV_CPU_CALL_AVX512_ICL(fn, args); __CV_EXPAND(__CV_CPU_DISPATCH_CHAIN_ ## mode(fn, args, __VA_ARGS__))
#if !defined CV_DISABLE_OPTIMIZATION && defined CV_ENABLE_INTRINSICS && defined CV_CPU_COMPILE_NEON
# define CV_TRY_NEON 1
# define CV_CPU_FORCE_NEON 1
......
modules/core/include/opencv2/core/cvdef.h
View file @ b2abd8ca
......@@ -235,6 +235,12 @@ namespace cv { namespace debug_build_guard { } using namespace debug_build_guard
#define CV_CPU_AVX_512PF 19
#define CV_CPU_AVX_512VBMI 20
#define CV_CPU_AVX_512VL 21
#define CV_CPU_AVX_512VBMI2 22
#define CV_CPU_AVX_512VNNI 23
#define CV_CPU_AVX_512BITALG 24
#define CV_CPU_AVX_512VPOPCNTDQ 25
#define CV_CPU_AVX_5124VNNIW 26
#define CV_CPU_AVX_5124FMAPS 27
#define CV_CPU_NEON 100
......@@ -243,6 +249,12 @@ namespace cv { namespace debug_build_guard { } using namespace debug_build_guard
// CPU features groups
#define CV_CPU_AVX512_SKX 256
#define CV_CPU_AVX512_COMMON 257
#define CV_CPU_AVX512_KNL 258
#define CV_CPU_AVX512_KNM 259
#define CV_CPU_AVX512_CNL 260
#define CV_CPU_AVX512_CEL 261
#define CV_CPU_AVX512_ICL 262
// when adding to this list remember to update the following enum
#define CV_HARDWARE_MAX_FEATURE 512
......@@ -273,6 +285,12 @@ enum CpuFeatures {
CPU_AVX_512PF = 19,
CPU_AVX_512VBMI = 20,
CPU_AVX_512VL = 21,
CPU_AVX_512VBMI2 = 22,
CPU_AVX_512VNNI = 23,
CPU_AVX_512BITALG = 24,
CPU_AVX_512VPOPCNTDQ= 25,
CPU_AVX_5124VNNIW = 26,
CPU_AVX_5124FMAPS = 27,
CPU_NEON = 100,
......@@ -280,6 +298,12 @@ enum CpuFeatures {
CPU_VSX3 = 201,
CPU_AVX512_SKX = 256, //!< Skylake-X with AVX-512F/CD/BW/DQ/VL
CPU_AVX512_COMMON = 257, //!< Common instructions AVX-512F/CD for all CPUs that support AVX-512
CPU_AVX512_KNL = 258, //!< Knights Landing with AVX-512F/CD/ER/PF
CPU_AVX512_KNM = 259, //!< Knights Mill with AVX-512F/CD/ER/PF/4FMAPS/4VNNIW/VPOPCNTDQ
CPU_AVX512_CNL = 260, //!< Cannon Lake with AVX-512F/CD/BW/DQ/VL/IFMA/VBMI
CPU_AVX512_CEL = 261, //!< Cascade Lake with AVX-512F/CD/BW/DQ/VL/IFMA/VBMI/VNNI
CPU_AVX512_ICL = 262, //!< Ice Lake with AVX-512F/CD/BW/DQ/VL/IFMA/VBMI/VNNI/VBMI2/BITALG/VPOPCNTDQ
CPU_MAX_FEATURE = 512 // see CV_HARDWARE_MAX_FEATURE
};
......
modules/core/src/system.cpp
View file @ b2abd8ca
......@@ -356,6 +356,12 @@ struct HWFeatures
g_hwFeatureNames[CPU_AVX_512PF] = "AVX512PF";
g_hwFeatureNames[CPU_AVX_512VBMI] = "AVX512VBMI";
g_hwFeatureNames[CPU_AVX_512VL] = "AVX512VL";
g_hwFeatureNames[CPU_AVX_512VBMI2] = "AVX512VBMI2";
g_hwFeatureNames[CPU_AVX_512VNNI] = "AVX512VNNI";
g_hwFeatureNames[CPU_AVX_512BITALG] = "AVX512BITALG";
g_hwFeatureNames[CPU_AVX_512VPOPCNTDQ] = "AVX512VPOPCNTDQ";
g_hwFeatureNames[CPU_AVX_5124VNNIW] = "AVX5124VNNIW";
g_hwFeatureNames[CPU_AVX_5124FMAPS] = "AVX5124FMAPS";
g_hwFeatureNames[CPU_NEON] = "NEON";
......@@ -363,6 +369,11 @@ struct HWFeatures
g_hwFeatureNames[CPU_VSX3] = "VSX3";
g_hwFeatureNames[CPU_AVX512_SKX] = "AVX512-SKX";
g_hwFeatureNames[CPU_AVX512_KNL] = "AVX512-KNL";
g_hwFeatureNames[CPU_AVX512_KNM] = "AVX512-KNM";
g_hwFeatureNames[CPU_AVX512_CNL] = "AVX512-CNL";
g_hwFeatureNames[CPU_AVX512_CEL] = "AVX512-CEL";
g_hwFeatureNames[CPU_AVX512_ICL] = "AVX512-ICL";
}
void initialize(void)
......@@ -404,15 +415,21 @@ struct HWFeatures
have[CV_CPU_AVX2] = (cpuid_data_ex[1] & (1<<5)) != 0;
have[CV_CPU_AVX_512F] = (cpuid_data_ex[1] & (1<<16)) != 0;
have[CV_CPU_AVX_512DQ] = (cpuid_data_ex[1] & (1<<17)) != 0;
have[CV_CPU_AVX_512IFMA512] = (cpuid_data_ex[1] & (1<<21)) != 0;
have[CV_CPU_AVX_512PF] = (cpuid_data_ex[1] & (1<<26)) != 0;
have[CV_CPU_AVX_512ER] = (cpuid_data_ex[1] & (1<<27)) != 0;
have[CV_CPU_AVX_512CD] = (cpuid_data_ex[1] & (1<<28)) != 0;
have[CV_CPU_AVX_512BW] = (cpuid_data_ex[1] & (1<<30)) != 0;
have[CV_CPU_AVX_512VL] = (cpuid_data_ex[1] & (1<<31)) != 0;
have[CV_CPU_AVX_512VBMI] = (cpuid_data_ex[2] & (1<<1)) != 0;
have[CV_CPU_AVX_512F] = (cpuid_data_ex[1] & (1<<16)) != 0;
have[CV_CPU_AVX_512DQ] = (cpuid_data_ex[1] & (1<<17)) != 0;
have[CV_CPU_AVX_512IFMA] = (cpuid_data_ex[1] & (1<<21)) != 0;
have[CV_CPU_AVX_512PF] = (cpuid_data_ex[1] & (1<<26)) != 0;
have[CV_CPU_AVX_512ER] = (cpuid_data_ex[1] & (1<<27)) != 0;
have[CV_CPU_AVX_512CD] = (cpuid_data_ex[1] & (1<<28)) != 0;
have[CV_CPU_AVX_512BW] = (cpuid_data_ex[1] & (1<<30)) != 0;
have[CV_CPU_AVX_512VL] = (cpuid_data_ex[1] & (1<<31)) != 0;
have[CV_CPU_AVX_512VBMI] = (cpuid_data_ex[2] & (1<<1)) != 0;
have[CV_CPU_AVX_512VBMI2] = (cpuid_data_ex[2] & (1<<6)) != 0;
have[CV_CPU_AVX_512VNNI] = (cpuid_data_ex[2] & (1<<11)) != 0;
have[CV_CPU_AVX_512BITALG] = (cpuid_data_ex[2] & (1<<12)) != 0;
have[CV_CPU_AVX_512VPOPCNTDQ] = (cpuid_data_ex[2] & (1<<14)) != 0;
have[CV_CPU_AVX_5124VNNIW] = (cpuid_data_ex[3] & (1<<2)) != 0;
have[CV_CPU_AVX_5124FMAPS] = (cpuid_data_ex[3] & (1<<3)) != 0;
bool have_AVX_OS_support = true;
bool have_AVX512_OS_support = true;
......@@ -446,15 +463,38 @@ struct HWFeatures
have[CV_CPU_AVX_512CD] = false;
have[CV_CPU_AVX_512DQ] = false;
have[CV_CPU_AVX_512ER] = false;
have[CV_CPU_AVX_512IFMA512] = false;
have[CV_CPU_AVX_512IFMA] = false;
have[CV_CPU_AVX_512PF] = false;
have[CV_CPU_AVX_512VBMI] = false;
have[CV_CPU_AVX_512VL] = false;
have[CV_CPU_AVX_512VBMI2] = false;
have[CV_CPU_AVX_512VNNI] = false;
have[CV_CPU_AVX_512BITALG] = false;
have[CV_CPU_AVX_512VPOPCNTDQ] = false;
have[CV_CPU_AVX_5124VNNIW] = false;
have[CV_CPU_AVX_5124FMAPS] = false;
}
if (have[CV_CPU_AVX_512F])
have[CV_CPU_AVX512_COMMON] = have[CV_CPU_AVX_512F] && have[CV_CPU_AVX_512CD];
if (have[CV_CPU_AVX512_COMMON])
{
have[CV_CPU_AVX512_SKX] = have[CV_CPU_AVX_512F] & have[CV_CPU_AVX_512CD] & have[CV_CPU_AVX_512BW] & have[CV_CPU_AVX_512DQ] & have[CV_CPU_AVX_512VL];
have[CV_CPU_AVX512_KNL] = have[CV_CPU_AVX_512ER] && have[CV_CPU_AVX_512PF];
have[CV_CPU_AVX512_KNM] = have[CV_CPU_AVX512_KNL] && have[CV_CPU_AVX_5124FMAPS] &&
have[CV_CPU_AVX_5124VNNIW] && have[CV_CPU_AVX_512VPOPCNTDQ];
have[CV_CPU_AVX512_SKX] = have[CV_CPU_AVX_512BW] && have[CV_CPU_AVX_512DQ] && have[CV_CPU_AVX_512VL];
have[CV_CPU_AVX512_CNL] = have[CV_CPU_AVX512_SKX] && have[CV_CPU_AVX_512IFMA] && have[CV_CPU_AVX_512VBMI];
have[CV_CPU_AVX512_CEL] = have[CV_CPU_AVX512_CNL] && have[CV_CPU_AVX_512VNNI];
have[CV_CPU_AVX512_ICL] = have[CV_CPU_AVX512_CEL] && have[CV_CPU_AVX_512VBMI2] &&
have[CV_CPU_AVX_512BITALG] && have[CV_CPU_AVX_512VPOPCNTDQ];
}
else
{
have[CV_CPU_AVX512_KNL] = false;
have[CV_CPU_AVX512_KNM] = false;
have[CV_CPU_AVX512_SKX] = false;
have[CV_CPU_AVX512_CNL] = false;
have[CV_CPU_AVX512_CEL] = false;
have[CV_CPU_AVX512_ICL] = false;
}
}
#endif // CV_CPUID_X86
......@@ -621,11 +661,14 @@ struct HWFeatures
}
if (isBaseline)
{
if (dump) fprintf(stderr, "OPENCV: Trying to disable baseline CPU feature: '%s'. This has very limited effect, because code optimizations for this feature are executed unconditionally in the most cases.\n", getHWFeatureNameSafe(i));
if (dump) fprintf(stderr, "OPENCV: Trying to disable baseline CPU feature: '%s'."
"This has very limited effect, because code optimizations for this feature are executed unconditionally "
"in the most cases.\n", getHWFeatureNameSafe(i));
}
if (!have[i])
{
if (dump) fprintf(stderr, "OPENCV: Trying to disable unavailable CPU feature on the current platform: '%s'.\n", getHWFeatureNameSafe(i));
if (dump) fprintf(stderr, "OPENCV: Trying to disable unavailable CPU feature on the current platform: '%s'.\n",
getHWFeatureNameSafe(i));
}
have[i] = false;
......
modules/dnn/src/layers/convolution_layer.cpp
View file @ b2abd8ca
......@@ -126,7 +126,7 @@ public:
inpShape.push_back(inputs[0].size[i]);
outShape.push_back(outputs[0].size[i]);
}
getConvPoolPaddings(inpShape, outShape, kernel_size, strides, padMode, dilations, pads_begin, pads_end);
getConvPoolPaddings(inpShape, kernel_size, strides, padMode, pads_begin, pads_end);
if (pads_begin.size() == 2) {
for (int i = 0; i < pads_begin.size(); i++) {
if (pads_begin[i] != pads_end[i])
......@@ -1331,7 +1331,7 @@ public:
inpShape.push_back(inputs[0].size[i]);
outShape.push_back(outputs[0].size[i]);
}
getConvPoolPaddings(outShape, inpShape, kernel_size, strides, padMode, dilations, pads_begin, pads_end);
getConvPoolPaddings(outShape, kernel_size, strides, padMode, pads_begin, pads_end);
if (pads_begin.size() == 2) {
for (int i = 0; i < pads_begin.size(); i++) {
if (pads_begin[i] != pads_end[i])
......
modules/dnn/src/layers/layers_common.cpp
View file @ b2abd8ca
......@@ -214,25 +214,25 @@ void getConvPoolOutParams(const std::vector<int>& inp, const std::vector<size_t>
}
}
void getConvPoolPaddings(const std::vector<int>& inp, const std::vector<int>& out,
const std::vector<size_t>& kernel, const std::vector<size_t>& strides,
const String &padMode, const std::vector<size_t>& dilation,
void getConvPoolPaddings(const std::vector<int>& inp, const std::vector<size_t>& kernel,
const std::vector<size_t>& strides, const String &padMode,
std::vector<size_t>& pads_begin, std::vector<size_t>& pads_end)
{
if (padMode == "VALID")
if (padMode == "SAME" || padMode == "VALID")
{
pads_begin.assign(kernel.size(), 0);
pads_end.assign(kernel.size(), 0);
}
else if (padMode == "SAME")
if (padMode == "SAME")
{
CV_Assert_N(kernel.size() == dilation.size(), kernel.size() == strides.size(),
kernel.size() == inp.size(), kernel.size() == out.size());
pads_begin.resize(kernel.size());
pads_end.resize(kernel.size());
CV_Assert_N(kernel.size() == strides.size(), kernel.size() == inp.size());
for (int i = 0; i < pads_begin.size(); i++) {
int pad = ((out[i] - 1) * strides[i] + dilation[i] * (kernel[i] - 1) + 1 - inp[i]) / 2;
pads_begin[i] = pads_end[i] = std::max(0, pad);
// There are test cases with stride > kernel.
if (strides[i] <= kernel[i])
{
int pad = (kernel[i] - 1 - (inp[i] - 1 + strides[i]) % strides[i]) / 2;
pads_begin[i] = pads_end[i] = pad;
}
}
}
}
......
modules/dnn/src/layers/layers_common.hpp
View file @ b2abd8ca
......@@ -69,9 +69,8 @@ void getConvPoolOutParams(const std::vector<int>& inp, const std::vector<size_t>
const std::vector<size_t>& stride, const String &padMode,
const std::vector<size_t>& dilation, std::vector<int>& out);
void getConvPoolPaddings(const std::vector<int>& inp, const std::vector<int>& out,
const std::vector<size_t>& kernel, const std::vector<size_t>& strides,
const String &padMode, const std::vector<size_t>& dilation,
void getConvPoolPaddings(const std::vector<int>& inp, const std::vector<size_t>& kernel,
const std::vector<size_t>& strides, const String &padMode,
std::vector<size_t>& pads_begin, std::vector<size_t>& pads_end);
}
}
......
modules/dnn/src/layers/pooling_layer.cpp
View file @ b2abd8ca
......@@ -144,7 +144,7 @@ public:
kernel_size = std::vector<size_t>(inp.begin(), inp.end());
}
getConvPoolPaddings(inp, out, kernel_size, strides, padMode, std::vector<size_t>(kernel_size.size(), 1), pads_begin, pads_end);
getConvPoolPaddings(inp, kernel_size, strides, padMode, pads_begin, pads_end);
if (pads_begin.size() == 2) {
pad_t = pads_begin[0];
pad_l = pads_begin[1];
......
modules/dnn/src/op_inf_engine.cpp
View file @ b2abd8ca
......@@ -958,7 +958,16 @@ Mat infEngineBlobToMat(const InferenceEngine::Blob::Ptr& blob)
// NOTE: Inference Engine sizes are reversed.
std::vector<size_t> dims = blob->dims();
std::vector<int> size(dims.rbegin(), dims.rend());
return Mat(size, CV_32F, (void*)blob->buffer());
int type = -1;
switch (blob->precision())
{
case InferenceEngine::Precision::FP32: type = CV_32F; break;
case InferenceEngine::Precision::U8: type = CV_8U; break;
default:
CV_Error(Error::StsNotImplemented, "Unsupported blob precision");
}
return Mat(size, type, (void*)blob->buffer());
}
bool InfEngineBackendLayer::getMemoryShapes(const std::vector<MatShape> &inputs,
......
modules/dnn/src/tensorflow/tf_graph_simplifier.cpp
View file @ b2abd8ca
......@@ -770,43 +770,47 @@ void RemoveIdentityOps(tensorflow::GraphDef& net)
}
}
Mat getTensorContent(const tensorflow::TensorProto &tensor)
Mat getTensorContent(const tensorflow::TensorProto &tensor, bool copy)
{
const std::string& content = tensor.tensor_content();
Mat m;
switch (tensor.dtype())
{
case tensorflow::DT_FLOAT:
{
if (!content.empty())
return Mat(1, content.size() / sizeof(float), CV_32FC1, (void*)content.c_str()).clone();
m = Mat(1, content.size() / sizeof(float), CV_32FC1, (void*)content.c_str());
else
{
const RepeatedField<float>& field = tensor.float_val();
CV_Assert(!field.empty());
return Mat(1, field.size(), CV_32FC1, (void*)field.data()).clone();
m = Mat(1, field.size(), CV_32FC1, (void*)field.data());
}
break;
}
case tensorflow::DT_DOUBLE:
{
if (!content.empty())
return Mat(1, content.size() / sizeof(double), CV_64FC1, (void*)content.c_str()).clone();
m = Mat(1, content.size() / sizeof(double), CV_64FC1, (void*)content.c_str());
else
{
const RepeatedField<double>& field = tensor.double_val();
CV_Assert(!field.empty());
return Mat(1, field.size(), CV_64FC1, (void*)field.data()).clone();
m = Mat(1, field.size(), CV_64FC1, (void*)field.data());
}
break;
}
case tensorflow::DT_INT32:
{
if (!content.empty())
return Mat(1, content.size() / sizeof(int32_t), CV_32SC1, (void*)content.c_str()).clone();
m = Mat(1, content.size() / sizeof(int32_t), CV_32SC1, (void*)content.c_str());
else
{
const RepeatedField<int32_t>& field = tensor.int_val();
CV_Assert(!field.empty());
return Mat(1, field.size(), CV_32SC1, (void*)field.data()).clone();
m = Mat(1, field.size(), CV_32SC1, (void*)field.data());
}
break;
}
case tensorflow::DT_HALF:
{
......@@ -825,20 +829,20 @@ Mat getTensorContent(const tensorflow::TensorProto &tensor)
}
// Reinterpret as a signed shorts just for a convertFp16 call.
Mat halfsSigned(halfs.size(), CV_16SC1, halfs.data);
Mat floats(halfs.size(), CV_32FC1);
convertFp16(halfsSigned, floats);
return floats;
convertFp16(halfsSigned, m);
break;
}
case tensorflow::DT_QUINT8:
{
CV_Assert(!content.empty());
return Mat(1, content.size(), CV_8UC1, (void*)content.c_str()).clone();
m = Mat(1, content.size(), CV_8UC1, (void*)content.c_str());
break;
}
default:
CV_Error(Error::StsError, "Tensor's data type is not supported");
break;
}
return Mat();
return copy ? m.clone() : m;
}
void releaseTensor(tensorflow::TensorProto* tensor)
......
modules/dnn/src/tensorflow/tf_graph_simplifier.hpp
View file @ b2abd8ca
......@@ -21,7 +21,7 @@ void RemoveIdentityOps(tensorflow::GraphDef& net);
void simplifySubgraphs(tensorflow::GraphDef& net);
Mat getTensorContent(const tensorflow::TensorProto &tensor);
Mat getTensorContent(const tensorflow::TensorProto &tensor, bool copy = true);
void releaseTensor(tensorflow::TensorProto* tensor);
......
modules/dnn/src/tensorflow/tf_importer.cpp
View file @ b2abd8ca
......@@ -109,7 +109,7 @@ void parseTensor(const tensorflow::TensorProto &tensor, Mat &dstBlob)
dstBlob.create(shape, CV_32F);
Mat tensorContent = getTensorContent(tensor);
Mat tensorContent = getTensorContent(tensor, /*no copy*/false);
int size = tensorContent.total();
CV_Assert(size == (int)dstBlob.total());
......@@ -509,7 +509,7 @@ void TFImporter::kernelFromTensor(const tensorflow::TensorProto &tensor, Mat &ds
dstBlob.create(shape, CV_32F);
Mat tensorContent = getTensorContent(tensor);
Mat tensorContent = getTensorContent(tensor, /*no copy*/false);
int size = tensorContent.total();
CV_Assert(size == (int)dstBlob.total());
......
modules/dnn/test/test_misc.cpp
View file @ b2abd8ca
......@@ -345,11 +345,12 @@ TEST(Net, forwardAndRetrieve)
#ifdef HAVE_INF_ENGINE
// This test runs network in synchronous mode for different inputs and then
// runs the same model asynchronously for the same inputs.
typedef testing::TestWithParam<Target> Async;
typedef testing::TestWithParam<tuple<int, Target> > Async;
TEST_P(Async, set_and_forward_single)
{
static const int kTimeout = 5000; // in milliseconds.
const int target = GetParam();
const int dtype = get<0>(GetParam());
const int target = get<1>(GetParam());
const std::string suffix = (target == DNN_TARGET_OPENCL_FP16 || target == DNN_TARGET_MYRIAD) ? "_fp16" : "";
const std::string& model = findDataFile("dnn/layers/layer_convolution" + suffix + ".bin");
......@@ -367,8 +368,8 @@ TEST_P(Async, set_and_forward_single)
int blobSize[] = {2, 6, 75, 113};
for (int i = 0; i < numInputs; ++i)
{
inputs[i].create(4, &blobSize[0], CV_32FC1);
randu(inputs[i], 0.0f, 1.0f);
inputs[i].create(4, &blobSize[0], dtype);
randu(inputs[i], 0, 255);
}
// Run synchronously.
......@@ -394,7 +395,8 @@ TEST_P(Async, set_and_forward_single)
TEST_P(Async, set_and_forward_all)
{
static const int kTimeout = 5000; // in milliseconds.
const int target = GetParam();
const int dtype = get<0>(GetParam());
const int target = get<1>(GetParam());
const std::string suffix = (target == DNN_TARGET_OPENCL_FP16 || target == DNN_TARGET_MYRIAD) ? "_fp16" : "";
const std::string& model = findDataFile("dnn/layers/layer_convolution" + suffix + ".bin");
......@@ -413,8 +415,8 @@ TEST_P(Async, set_and_forward_all)
int blobSize[] = {2, 6, 75, 113};
for (int i = 0; i < numInputs; ++i)
{
inputs[i].create(4, &blobSize[0], CV_32FC1);
randu(inputs[i], 0.0f, 1.0f);
inputs[i].create(4, &blobSize[0], dtype);
randu(inputs[i], 0, 255);
}
// Run synchronously.
......@@ -441,7 +443,10 @@ TEST_P(Async, set_and_forward_all)
}
}
INSTANTIATE_TEST_CASE_P(/**/, Async, testing::ValuesIn(getAvailableTargets(DNN_BACKEND_INFERENCE_ENGINE)));
INSTANTIATE_TEST_CASE_P(/**/, Async, Combine(
Values(CV_32F, CV_8U),
testing::ValuesIn(getAvailableTargets(DNN_BACKEND_INFERENCE_ENGINE))
));
#endif // HAVE_INF_ENGINE
}} // namespace
modules/ml/doc/ml_intro.markdown
View file @ b2abd8ca
......@@ -25,7 +25,7 @@ components:
vector responses.
- Another optional component is the mask of missing measurements. Most algorithms require all the
components in all the training samples be valid, but some other algorithms, such as decision
tress, can handle the cases of missing measurements.
trees, can handle the cases of missing measurements.
- In the case of classification problem user may want to give different weights to different
classes. This is useful, for example, when:
- user wants to shift prediction accuracy towards lower false-alarm rate or higher hit-rate.
......
samples/dnn/tf_text_graph_ssd.py
View file @ b2abd8ca
......@@ -274,7 +274,8 @@ def createSSDGraph(modelPath, configPath, outputPath):
num_matched_layers = 0
for node in graph_def.node:
if re.match('BoxPredictor_\d/BoxEncodingPredictor/Conv2D', node.name) or \
if re.match('BoxPredictor_\d/BoxEncodingPredictor/convolution', node.name) or \
re.match('BoxPredictor_\d/BoxEncodingPredictor/Conv2D', node.name) or \
re.match('WeightSharedConvolutionalBoxPredictor(_\d)*/BoxPredictor/Conv2D', node.name):
node.addAttr('loc_pred_transposed', True)
num_matched_layers += 1
......
Markdown is supported
0% or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment