[MaskRCNN] small updates on docs

examples/FasterRCNN/config.py

@@ -63,7 +63,7 @@ CROWD_OVERLAP_THRES = 0.7  # boxes overlapping crowd will be ignored.
 
 # fastrcnn training ---------------------
 FASTRCNN_BATCH_PER_IM = 512
-FASTRCNN_BBOX_REG_WEIGHTS = np.array([10, 10, 5, 5], dtype='float32')
+FASTRCNN_BBOX_REG_WEIGHTS = [10., 10., 5., 5.]  # Better but non-standard setting: [20, 20, 10, 10]
 FASTRCNN_FG_THRESH = 0.5
 FASTRCNN_FG_RATIO = 0.25  # fg ratio in a ROI batch
 

examples/FasterRCNN/data.py

@@ -154,8 +154,7 @@ def get_anchor_labels(anchors, gt_boxes, crowd_boxes):
     # Subsample bg labels. num_bg is not allowed to be too many
     old_num_bg = np.sum(anchor_labels == 0)
     if old_num_bg == 0:
-        # No valid bg/fg in this image, skip.
-        # This can happen if, e.g. the image has large crowd.
+        # No valid bg in this image, skip.
         raise MalformedData("No valid background for RPN!")
     target_num_bg = config.RPN_BATCH_PER_IM - len(fg_inds)
     filter_box_label(anchor_labels, 0, target_num_bg)   # ignore return values

examples/FasterRCNN/model.py

@@ -306,7 +306,7 @@ def sample_fast_rcnn_targets(boxes, gt_boxes, gt_labels):
     ret_labels = tf.concat(
         [tf.gather(gt_labels, fg_inds_wrt_gt),
          tf.zeros_like(bg_inds, dtype=tf.int64)], axis=0)
-    # stop the gradient -- they are meant to be ground-truth
+    # stop the gradient -- they are meant to be training targets
     return tf.stop_gradient(ret_boxes, name='sampled_proposal_boxes'), \
         tf.stop_gradient(ret_labels, name='sampled_labels'), \
         tf.stop_gradient(fg_inds_wrt_gt)

examples/FasterRCNN/train.py

@@ -76,8 +76,9 @@ class DetectionModel(ModelDesc):
     @under_name_scope()
     def narrow_to_featuremap(self, featuremap, anchors, anchor_labels, anchor_boxes):
         """
-        Args:
         Slice anchors/anchor_labels/anchor_boxes to the spatial size of this featuremap.
+
+        Args:
             anchors (FS x FS x NA x 4):
             anchor_labels (FS x FS x NA):
             anchor_boxes (FS x FS x NA x 4):
@@ -112,7 +113,7 @@ class DetectionModel(ModelDesc):
             fg_rcnn_boxes (fg x 4): proposal boxes for each sampled foreground targets
             gt_boxes_per_fg (fg x 4): matching gt boxes for each sampled foreground targets
             rcnn_label_logits (n): label logits for each sampled targets
-            fg_rcnn_box_logits (fg x 4): box logits for each sampled foreground targets
+            fg_rcnn_box_logits (fg x #class x 4): box logits for each sampled foreground targets
         """
 
         with tf.name_scope('fg_sample_patch_viz'):
@@ -124,7 +125,7 @@ class DetectionModel(ModelDesc):
             tf.summary.image('viz', fg_sampled_patches, max_outputs=30)
 
         encoded_boxes = encode_bbox_target(
-            gt_boxes_per_fg, fg_rcnn_boxes) * tf.constant(config.FASTRCNN_BBOX_REG_WEIGHTS)
+            gt_boxes_per_fg, fg_rcnn_boxes) * tf.constant(config.FASTRCNN_BBOX_REG_WEIGHTS, dtype=tf.float32)
         fastrcnn_label_loss, fastrcnn_box_loss = fastrcnn_losses(
             rcnn_labels, rcnn_label_logits,
             encoded_boxes,
@@ -138,7 +139,7 @@ class DetectionModel(ModelDesc):
             image_shape2d: h, w
             rcnn_boxes (nx4): the proposal boxes
             rcnn_label_logits (n):
-            rcnn_box_logits (nx4):
+            rcnn_box_logits (nx #class x 4):
 
         Returns:
             boxes (mx4):
@@ -148,7 +149,7 @@ class DetectionModel(ModelDesc):
         anchors = tf.tile(tf.expand_dims(rcnn_boxes, 1), [1, config.NUM_CLASS - 1, 1])   # #proposal x #Cat x 4
         decoded_boxes = decode_bbox_target(
             rcnn_box_logits /
-            tf.constant(config.FASTRCNN_BBOX_REG_WEIGHTS), anchors)
+            tf.constant(config.FASTRCNN_BBOX_REG_WEIGHTS, dtype=tf.float32), anchors)
         decoded_boxes = clip_boxes(decoded_boxes, image_shape2d, name='fastrcnn_all_boxes')
 
         # indices: Nx2. Each index into (#proposal, #category)
@@ -332,7 +333,8 @@ class ResNetFPNModel(DetectionModel):
         c2345 = resnet_fpn_backbone(image, config.RESNET_NUM_BLOCK)
         p23456 = fpn_model('fpn', c2345)
 
-        # images are padded for p5, which are too large for p2-p4
+        # Images are padded for p5, which are too large for p2-p4.
+        # This seems to have no effect on mAP.
         for i, stride in enumerate(config.ANCHOR_STRIDES_FPN[:3]):
             pi = p23456[i]
             target_shape = tf.to_int32(tf.ceil(tf.to_float(image_shape2d) * (1.0 / stride)))
@@ -423,7 +425,7 @@ class ResNetFPNModel(DetectionModel):
                 mrcnn_loss = 0.0
 
             wd_cost = regularize_cost(
-                '(?:group1|group2|group3|rpn|fastrcnn|maskrcnn)/.*W',
+                '(?:group1|group2|group3|rpn|fpn|fastrcnn|maskrcnn)/.*W',
                 l2_regularizer(1e-4), name='wd_cost')
 
             total_cost = tf.add_n(rpn_loss_collection + [