[FastRCNN] fix box selection in evaluation

examples/FasterRCNN/README.md

@@ -46,7 +46,7 @@ To evaluate the performance (pretrained models can be downloaded in [model zoo](
 
 Mean Average Precision @IoU=0.50:0.95:
 
-+ trainval35k/minival, FASTRCNN_BATCH=256: 33.4. Takes 49h on 8 TitanX.
++ trainval35k/minival, FASTRCNN_BATCH=256: 33.7. Takes 49h on 8 TitanX.
 + trainval35k/minival, FASTRCNN_BATCH=64: 32.2. Takes 31h on 8 TitanX.
 
 The hyperparameters are not carefully tuned. You can probably get better performance by e.g.  training longer.

examples/FasterRCNN/eval.py

@@ -44,8 +44,8 @@ def get_tf_nms(num_output, thresh):
 def nms_fastrcnn_results(boxes, probs):
     """
     Args:
-        boxes: nx4 floatbox in float32
-        probs: nxC
+        boxes: nx#catx4 floatbox in float32
+        probs: nx#class
 
     Returns:
         [DetectionResult]
@@ -60,7 +60,7 @@ def nms_fastrcnn_results(boxes, probs):
         if ids.size == 0:
             continue
         probs_k = probs[ids, klass].flatten()
-        boxes_k = boxes[ids, :]
+        boxes_k = boxes[ids, klass - 1, :]
         selected_ids = nms_func(boxes_k, probs_k)
         selected_boxes = boxes_k[selected_ids, :].copy()
         ret.append(DetectionResult(klass, selected_boxes, probs_k[selected_ids]))

examples/FasterRCNN/train.py

@@ -137,14 +137,11 @@ class Model(ModelDesc):
                 add_moving_summary(k)
         else:
             label_probs = tf.nn.softmax(fastrcnn_label_logits, name='fastrcnn_all_probs')  # #proposal x #Class
-            labels = tf.argmax(fastrcnn_label_logits, axis=1)
-            fg_ind, fg_box_logits = fastrcnn_predict_boxes(labels, fastrcnn_box_logits)
-            fg_label_probs = tf.gather(label_probs, fg_ind, name='fastrcnn_fg_probs')
-            fg_boxes = tf.gather(proposal_boxes, fg_ind)
-
-            fg_box_logits = fg_box_logits / tf.constant(config.FASTRCNN_BBOX_REG_WEIGHTS)
-            decoded_boxes = decode_bbox_target(fg_box_logits, fg_boxes)  # #fgx4, floatbox
-            decoded_boxes = tf.identity(decoded_boxes, name='fastrcnn_fg_boxes')
+            anchors = tf.tile(tf.expand_dims(proposal_boxes, 1), [1, config.NUM_CLASS - 1, 1])   # #proposal x #Cat x 4
+            decoded_boxes = decode_bbox_target(
+                fastrcnn_box_logits /
+                tf.constant(config.FASTRCNN_BBOX_REG_WEIGHTS), anchors)
+            decoded_boxes = tf.identity(decoded_boxes, name='fastrcnn_all_boxes')
 
     def _get_optimizer(self):
         lr = tf.get_variable('learning_rate', initializer=0.003, trainable=False)
@@ -210,8 +207,8 @@ def offline_evaluate(model_path, output_file):
         session_init=get_model_loader(model_path),
         input_names=['image'],
         output_names=[
-            'fastrcnn_fg_probs',
-            'fastrcnn_fg_boxes',
+            'fastrcnn_all_probs',
+            'fastrcnn_all_boxes',
         ]))
     df = get_eval_dataflow()
     df = PrefetchDataZMQ(df, 1)
@@ -227,8 +224,8 @@ def predict(model_path, input_file):
         session_init=get_model_loader(model_path),
         input_names=['image'],
         output_names=[
-            'fastrcnn_fg_probs',
-            'fastrcnn_fg_boxes',
+            'fastrcnn_all_probs',
+            'fastrcnn_all_boxes',
         ]))
     img = cv2.imread(input_file, cv2.IMREAD_COLOR)
     results = detect_one_image(img, pred)
@@ -239,7 +236,8 @@ def predict(model_path, input_file):
 
 class EvalCallback(Callback):
     def _setup_graph(self):
-        self.pred = self.trainer.get_predictor(['image'], ['fastrcnn_fg_probs', 'fastrcnn_fg_boxes'])
+        self.pred = self.trainer.get_predictor(
+            ['image'], ['fastrcnn_all_probs', 'fastrcnn_all_boxes'])
         self.df = PrefetchDataZMQ(get_eval_dataflow(), 1)
 
     def _before_train(self):

examples/README.md

@@ -43,7 +43,6 @@ Training examples with __reproducible__ and meaningful performance.
 
 Example needs to satisfy one of the following:
 + Reproduce performance of a published or well-known paper.
-+ Get state-of-the-art performance on some task.
 + Illustrate a new way of using the library that is currently not covered.
 
 __Performance is important__. Usually deep learning code is easy to write,