FPN results

README.md

@@ -3,16 +3,16 @@ Feature Pyramid Network on caffe
 This is the unoffical version  Feature Pyramid Network for Feature Pyramid Networks for Object Detection https://arxiv.org/abs/1612.03144
 
 # results
-`FPN(resnet50) result is implemented without OHEM and train with pascal voc 2007 + 2012 test on 2007`
+`FPN(resnet50)-end2end result is implemented without OHEM and train with pascal voc 2007 + 2012 test on 2007`
 
 
 |[email protected]|aeroplane|bicycle|bird|boat|bottle|bus|car|cat|chair|cow|
 |:--:|:-------:| -----:| --:| --:|-----:|--:|--:|--:|----:|--:|
-
+|0.7833|0.8585| 0.8001| 0.7970| 0.7174|0.6522|0.8668|0.8768|0.8929|0.5842|0.8658|
 
 |diningtable|dog |horse|motorbike|person |pottedplant|sheep|sofa|train|tv|
 |----------:|:--:|:---:| -------:| -----:| -------:|----:|---:|----:|--:|
-
+|0.7022|0.8891|0.8680| 0.7991| 0.7944| 0.5065|0.7896|0.7707|0.8697|0.7653|
 # framework
 ![](framework.png)
 `the red and yellow are shared params`

lib/fast_rcnn/config.py

@@ -170,6 +170,9 @@
 __C.TEST.RPN_POST_NMS_TOP_N = 1000
 # Proposal height and width both need to be greater than RPN_MIN_SIZE (at orig image scale)
 __C.TEST.RPN_MIN_SIZE = 16
+# Apply box scoring heuristics
+__C.TEST.BBOX_VOTE_N_WEIGHTED_SCORE = 1
+__C.TEST.BBOX_VOTE_WEIGHT_EMPTY = 0.5
 
 
 #

lib/fast_rcnn/test.py

@@ -18,6 +18,7 @@
 import cPickle
 from utils.blob import im_list_to_blob
 import os
+from utils.cython_bbox import bbox_overlaps
 
 def _get_image_blob(im):
     """Converts an image into a network input.
@@ -174,6 +175,42 @@ def vis_rois_detection(im_array, detections):
     plt.cla()
 
     plt. close(0)
+def bbox_vote(dets_NMS, dets_all, thresh=0.5):
+    dets_voted = np.zeros_like(dets_NMS)   # Empty matrix with the same shape and type
+
+    _overlaps = bbox_overlaps(
+			np.ascontiguousarray(dets_NMS[:, 0:4], dtype=np.float),
+			np.ascontiguousarray(dets_all[:, 0:4], dtype=np.float))
+
+    # for each survived box
+    for i, det in enumerate(dets_NMS):
+        dets_overlapped = dets_all[np.where(_overlaps[i, :] >= thresh)[0]]
+        assert(len(dets_overlapped) > 0)
+
+        boxes = dets_overlapped[:, 0:4]
+        scores = dets_overlapped[:, 4]
+
+        out_box = np.dot(scores, boxes)
+
+        dets_voted[i][0:4] = out_box / sum(scores)        # Weighted bounding boxes
+        dets_voted[i][4] = det[4]                         # Keep the original score
+
+        # Weighted scores (if enabled)
+        if cfg.TEST.BBOX_VOTE_N_WEIGHTED_SCORE > 1:
+            n_agreement = cfg.TEST.BBOX_VOTE_N_WEIGHTED_SCORE
+            w_empty = cfg.TEST.BBOX_VOTE_WEIGHT_EMPTY
+
+            n_detected = len(scores)
+
+            if n_detected >= n_agreement:
+                top_scores = -np.sort(-scores)[:n_agreement]
+                new_score = np.average(top_scores)
+            else:
+                new_score = np.average(scores) * (n_detected * 1.0 + (n_agreement - n_detected) * w_empty) / n_agreement
+
+            dets_voted[i][4] = min(new_score, dets_voted[i][4])
+
+    return dets_voted
 
 def im_detect(net, im, boxes=None,num_classes=21):
     """Detect object classes in an image given object proposals.
@@ -404,7 +441,8 @@ def test_net(net, imdb, max_per_image=100, thresh=0.05, vis=False):
                 .astype(np.float32, copy=False)
             keep = nms(cls_dets, cfg.TEST.NMS)
 
-            cls_dets = cls_dets[keep, :]
+            dets_NMSed = cls_dets[keep, :]
+            cls_dets = bbox_vote(dets_NMSed, cls_dets)
            # print cls_scores
             if vis:
                 vis_detections(im, imdb.classes[j], cls_dets)