Add clipping before visulization in detection model. (#1415)

* Fix label lists for inference.
* Add clipping before visulization.

fluid/PaddleCV/object_detection/README_quant.md

@@ -43,7 +43,7 @@ A Python transpiler is used to rewrite Fluid training program or evaluation prog
 
   The training is fine-tuned on the well-trained MobileNet-SSD model. So download model at first:
 
-  ```bash
+  ```
   wget http://paddlemodels.bj.bcebos.com/ssd_mobilenet_v1_pascalvoc.tar.gz
   ```
 
@@ -111,7 +111,7 @@ A Python transpiler is used to rewrite Fluid training program or evaluation prog
 
   Users can evaluate the converted model by:
 
-  ```bash
+  ```
   python main_quant.py \
       --data_dir=$PascalVOC_DIR$ \
       --mode='test' \
@@ -121,10 +121,11 @@ A Python transpiler is used to rewrite Fluid training program or evaluation prog
 
   You also can check the 8-bit model by the inference scripts
 
-  ```bash
+  ```
   python main_quant.py \
       --mode='infer' \
       --init_model=$MobileNet_SSD_8BIT_MODEL$ \
+      --confs_threshold=0.5 \
       --image_path='/data/PascalVOC/VOCdevkit/VOC2007/JPEGImages/002271.jpg'
   ```
   See 002271.jpg for the visualized image with bbouding boxes.

fluid/PaddleCV/object_detection/infer.py

@@ -85,11 +85,11 @@ def draw_bounding_box_on_image(image_path, nms_out, confs_threshold,
     im_width, im_height = image.size
 
     for dt in nms_out:
-        category_id, score, xmin, ymin, xmax, ymax = dt.tolist()
-        if score < confs_threshold:
+        if dt[1] < confs_threshold:
             continue
+        category_id = dt[0]
         bbox = dt[2:]
-        xmin, ymin, xmax, ymax = bbox
+        xmin, ymin, xmax, ymax = clip_bbox(dt[2:])
         (left, right, top, bottom) = (xmin * im_width, xmax * im_width,
                                       ymin * im_height, ymax * im_height)
         draw.line(
@@ -104,6 +104,14 @@ def draw_bounding_box_on_image(image_path, nms_out, confs_threshold,
     image.save(image_name)
 
 
+def clip_bbox(bbox):
+    xmin = max(min(bbox[0], 1.), 0.)
+    ymin = max(min(bbox[1], 1.), 0.)
+    xmax = max(min(bbox[2], 1.), 0.)
+    ymax = max(min(bbox[3], 1.), 0.)
+    return xmin, ymin, xmax, ymax
+
+
 if __name__ == '__main__':
     args = parser.parse_args()
     print_arguments(args)

fluid/PaddleCV/object_detection/main_quant.py

@@ -227,8 +227,8 @@ def infer(args, data_args):
         executor=exe,
         model_filename='__model__')
 
-    print(np.array(fluid.global_scope().find_var('conv2d_20.w_0').get_tensor()))
-    print(np.max(np.array(fluid.global_scope().find_var('conv2d_20.w_0').get_tensor())))
+    #print(np.array(fluid.global_scope().find_var('conv2d_20.w_0').get_tensor()))
+    #print(np.max(np.array(fluid.global_scope().find_var('conv2d_20.w_0').get_tensor())))
     infer_reader = reader.infer(data_args, image_path)
     data = infer_reader()
     data = data.reshape((1,) + data.shape)