Fixes for OpenCV face detection network

dkurt · dkurt · commit f16c7ee848f9 · 2019-07-27T15:09:15.000+03:00
diff --git a/modules/dnn/misc/face_detector_accuracy.py b/modules/dnn/misc/face_detector_accuracy.py
@@ -154,8 +154,13 @@ def detect(img, imageId):
             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)
diff --git a/modules/dnn/misc/quantize_face_detector.py b/modules/dnn/misc/quantize_face_detector.py
@@ -173,9 +173,9 @@ def l2norm(x, name):
 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')
 
diff --git a/modules/dnn/src/layers/detection_output_layer.cpp b/modules/dnn/src/layers/detection_output_layer.cpp
@@ -198,7 +198,7 @@ class DetectionOutputLayerImpl CV_FINAL : public DetectionOutputLayer
     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 @@ class DetectionOutputLayerImpl CV_FINAL : public DetectionOutputLayer
 
         InferenceEngine::Builder::Layer l = ieLayer;
         l.getParameters()["eta"] = std::string("1.0");
+        l.getParameters()["clip"] = _clip;
 
         return Ptr<BackendNode>(new InfEngineBackendNode(l));
     }
diff --git a/samples/dnn/face_detector/deploy.prototxt b/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
   }
 }
diff --git a/samples/dnn/face_detector/deploy_lowres.prototxt b/samples/dnn/face_detector/deploy_lowres.prototxt
@@ -1786,5 +1786,6 @@ layer {
     code_type: CENTER_SIZE
     keep_top_k: 200
     confidence_threshold: 0.01
+    clip: 1
   }
 }
diff --git a/samples/dnn/face_detector/opencv_face_detector.pbtxt b/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"
diff --git a/samples/dnn/face_detector/test.prototxt b/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 {
diff --git a/samples/dnn/face_detector/train.prototxt b/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 {

Original file line number	Diff line number	Diff line change
`@@ -198,7 +198,7 @@ class DetectionOutputLayerImpl CV_FINAL : public DetectionOutputLayer`
`198`	`198`	`virtual bool supportBackend(int backendId) CV_OVERRIDE`
`199`	`199`	`{`
`200`	`200`	`return backendId == DNN_BACKEND_OPENCV \|\|`
`201`		`- (backendId == DNN_BACKEND_INFERENCE_ENGINE && !_locPredTransposed && _bboxesNormalized && !_clip);`
	`201`	`+ (backendId == DNN_BACKEND_INFERENCE_ENGINE && !_locPredTransposed && _bboxesNormalized);`
`202`	`202`	`}`
`203`	`203`
`204`	`204`	`bool getMemoryShapes(const std::vector<MatShape> &inputs,`
`@@ -936,6 +936,7 @@ class DetectionOutputLayerImpl CV_FINAL : public DetectionOutputLayer`
`936`	`936`
`937`	`937`	`InferenceEngine::Builder::Layer l = ieLayer;`
`938`	`938`	`l.getParameters()["eta"] = std::string("1.0");`
	`939`	`+ l.getParameters()["clip"] = _clip;`
`939`	`940`
`940`	`941`	`return Ptr<BackendNode>(new InfEngineBackendNode(l));`
`941`	`942`	`}`
Original file line number	Diff line number	Diff line change
`@@ -1784,7 +1784,7 @@ layer {`
`1784`	`1784`	`}`
`1785`	`1785`	`code_type: CENTER_SIZE`
`1786`	`1786`	`keep_top_k: 200`
`1787`		`- confidence_threshold: 0.2`
	`1787`	`+ confidence_threshold: 0.01`
`1788`	`1788`	`clip: 1`
`1789`	`1789`	`}`
`1790`	`1790`	`}`
Original file line number	Diff line number	Diff line change
`@@ -1786,5 +1786,6 @@ layer {`
`1786`	`1786`	`code_type: CENTER_SIZE`
`1787`	`1787`	`keep_top_k: 200`
`1788`	`1788`	`confidence_threshold: 0.01`
	`1789`	`+ clip: 1`
`1789`	`1790`	`}`
`1790`	`1791`	`}`
Original file line number	Diff line number	Diff line change
`@@ -1221,7 +1221,7 @@ node {`
`1221`	`1221`	`attr {`
`1222`	`1222`	`key: "padding"`
`1223`	`1223`	`value {`
`1224`		`- s: "SAME"`
	`1224`	`+ s: "VALID"`
`1225`	`1225`	`}`
`1226`	`1226`	`}`
`1227`	`1227`	`attr {`
`@@ -1311,7 +1311,7 @@ node {`
`1311`	`1311`	`attr {`
`1312`	`1312`	`key: "padding"`
`1313`	`1313`	`value {`
`1314`		`- s: "SAME"`
	`1314`	`+ s: "VALID"`
`1315`	`1315`	`}`
`1316`	`1316`	`}`
`1317`	`1317`	`attr {`
`@@ -2337,6 +2337,12 @@ node {`
`2337`	`2337`	`i: 400`
`2338`	`2338`	`}`
`2339`	`2339`	`}`
	`2340`	`+ attr {`
	`2341`	`+ key: "clip"`
	`2342`	`+ value {`
	`2343`	`+ b: true`
	`2344`	`+ }`
	`2345`	`+ }`
`2340`	`2346`	`}`
`2341`	`2347`	`node {`
`2342`	`2348`	`name: "reshape_before_softmax"`
Original file line number	Diff line number	Diff line change
`@@ -917,7 +917,7 @@ layer {`
`917`	`917`	`}`
`918`	`918`	`convolution_param {`
`919`	`919`	`num_output: 128`
`920`		`- pad: 1`
	`920`	`+ pad: 0`
`921`	`921`	`kernel_size: 3`
`922`	`922`	`stride: 1`
`923`	`923`	`weight_filler {`
`@@ -983,7 +983,7 @@ layer {`
`983`	`983`	`}`
`984`	`984`	`convolution_param {`
`985`	`985`	`num_output: 128`
`986`		`- pad: 1`
	`986`	`+ pad: 0`
`987`	`987`	`kernel_size: 3`
`988`	`988`	`stride: 1`
`989`	`989`	`weight_filler {`
`@@ -1810,6 +1810,7 @@ layer {`
`1810`	`1810`	`code_type: CENTER_SIZE`
`1811`	`1811`	`keep_top_k: 200`
`1812`	`1812`	`confidence_threshold: 0.01`
	`1813`	`+ clip: 1`
`1813`	`1814`	`}`
`1814`	`1815`	`}`
`1815`	`1816`	`layer {`