Fixes for OpenCV face detection network

modules/dnn/misc/face_detector_accuracy.py

@@ -154,8 +154,13 @@ if args.proto and args.model:
             top = int(out[0, 0, i, 4] * img.shape[0])
             right = int(out[0, 0, i, 5] * img.shape[1])
             bottom = int(out[0, 0, i, 6] * img.shape[0])
-            addDetection(detections, imageId, left, top, width=right - left + 1,
-                         height=bottom - top + 1, score=confidence)
+
+            x = max(0, min(left, img.shape[1] - 1))
+            y = max(0, min(top, img.shape[0] - 1))
+            w = max(0, min(right - x + 1, img.shape[1] - x))
+            h = max(0, min(bottom - y + 1, img.shape[0] - y))
+
+            addDetection(detections, imageId, x, y, w, h, score=confidence)
 
 elif args.cascade:
     cascade = cv.CascadeClassifier(args.cascade)

modules/dnn/misc/quantize_face_detector.py

@@ -173,9 +173,9 @@ conv7_2_h = tf.space_to_batch_nd(conv7_1_h, [1, 1], [[1, 1], [1, 1]], name='Pad_
 conv7_2_h = conv(conv7_2_h, stride=2, pad='VALID', name='conv7_2_h', activ=tf.nn.relu)
 
 conv8_1_h = conv(conv7_2_h, pad='SAME', name='conv8_1_h', activ=tf.nn.relu)
-conv8_2_h = conv(conv8_1_h, pad='SAME', name='conv8_2_h', activ=tf.nn.relu)
+conv8_2_h = conv(conv8_1_h, pad='VALID', name='conv8_2_h', activ=tf.nn.relu)
 conv9_1_h = conv(conv8_2_h, 'conv9_1_h', activ=tf.nn.relu)
-conv9_2_h = conv(conv9_1_h, pad='SAME', name='conv9_2_h', activ=tf.nn.relu)
+conv9_2_h = conv(conv9_1_h, pad='VALID', name='conv9_2_h', activ=tf.nn.relu)
 
 conv4_3_norm = l2norm(layer_256_1_relu1, 'conv4_3_norm')
 

modules/dnn/src/layers/detection_output_layer.cpp

@@ -198,7 +198,7 @@ public:
     virtual bool supportBackend(int backendId) CV_OVERRIDE
     {
         return backendId == DNN_BACKEND_OPENCV ||
-               (backendId == DNN_BACKEND_INFERENCE_ENGINE && !_locPredTransposed && _bboxesNormalized && !_clip);
+               (backendId == DNN_BACKEND_INFERENCE_ENGINE && !_locPredTransposed && _bboxesNormalized);
     }
 
     bool getMemoryShapes(const std::vector<MatShape> &inputs,
@@ -936,6 +936,7 @@ public:
 
         InferenceEngine::Builder::Layer l = ieLayer;
         l.getParameters()["eta"] = std::string("1.0");
+        l.getParameters()["clip"] = _clip;
 
         return Ptr<BackendNode>(new InfEngineBackendNode(l));
     }

samples/dnn/face_detector/deploy.prototxt

@@ -1784,7 +1784,7 @@ layer {
     }
     code_type: CENTER_SIZE
     keep_top_k: 200
-    confidence_threshold: 0.2
+    confidence_threshold: 0.01
     clip: 1
   }
 }

samples/dnn/face_detector/opencv_face_detector.pbtxt

@@ -1221,7 +1221,7 @@ node {
   attr {
     key: "padding"
     value {
-      s: "SAME"
+      s: "VALID"
     }
   }
   attr {
@@ -1311,7 +1311,7 @@ node {
   attr {
     key: "padding"
     value {
-      s: "SAME"
+      s: "VALID"
     }
   }
   attr {
@@ -2337,6 +2337,12 @@ node {
       i: 400
     }
   }
+  attr {
+    key: "clip"
+    value {
+      b: true
+    }
+  }
 }
 node {
   name: "reshape_before_softmax"

samples/dnn/face_detector/test.prototxt

@@ -917,7 +917,7 @@ layer {
   }
   convolution_param {
     num_output: 128
-    pad: 1
+    pad: 0
     kernel_size: 3
     stride: 1
     weight_filler {
@@ -983,7 +983,7 @@ layer {
   }
   convolution_param {
     num_output: 128
-    pad: 1
+    pad: 0
     kernel_size: 3
     stride: 1
     weight_filler {
@@ -1810,6 +1810,7 @@ layer {
     code_type: CENTER_SIZE
     keep_top_k: 200
     confidence_threshold: 0.01
+    clip: 1
   }
 }
 layer {

samples/dnn/face_detector/train.prototxt

@@ -1086,7 +1086,7 @@ layer {
   }
   convolution_param {
     num_output: 128
-    pad: 1
+    pad: 0
     kernel_size: 3
     stride: 1
     weight_filler {
@@ -1600,7 +1600,7 @@ layer {
   }
   convolution_param {
     num_output: 16
-    pad: 1
+    pad: 0
     kernel_size: 3
     stride: 1
     weight_filler {