TEST: BINARY -> SVM

experiments/svm.yml

 EXP_DIR: svm
+TRAIN:
+  # don't use flipped examples when training SVMs for two reasons:
+  # 1) R-CNN didn't
+  # 2) I've tried and it doesn't help, yet makes SVM training take 2x longer
+  USE_FLIPPED: False
 TEST:
-  BINARY: True
+  SVM: True

src/fast_rcnn_config.py

@@ -97,9 +97,9 @@ __C.TEST.MAX_SIZE   = 1000
 # IoU >= this threshold)
 __C.TEST.NMS        = 0.3
 
-# Experimental: use binary logistic regression scores instead of K-way softmax
-# scores when testing
-__C.TEST.BINARY     = False
+# Experimental: treat the (K+1) units in the cls_score layer as linear
+# predictors (trained, eg, with one-vs-rest SVMs).
+__C.TEST.SVM             = False
 
 # Test using bounding-box regressors
 __C.TEST.BBOX_REG        = True

src/fast_rcnn_test.py

@@ -136,11 +136,10 @@ def im_detect(net, im, boxes):
     net.blobs['rois'].reshape(*(blobs['rois'].shape))
     blobs_out = net.forward(data=blobs['data'].astype(np.float32, copy=False),
                             rois=blobs['rois'].astype(np.float32, copy=False))
-    if cfg.TEST.BINARY:
-        # simulate binary logistic regression
+    if cfg.TEST.SVM:
+        # use the raw scores before softmax under the assumption they
+        # were trained as linear SVMs
         scores = net.blobs['cls_score'].data
-        # Return scores as fg - bg
-        scores = scores - scores[:, 0][:, np.newaxis]
     else:
         # use softmax estimated probabilities
         scores = blobs_out['cls_prob']

tools/extra/train_svms.py

@@ -265,10 +265,12 @@ def parse_args():
 
 if __name__ == '__main__':
     # Must turn this off to prevent issues when digging into the net blobs to
-    # pull out features
+    # pull out features (tricky!)
     cfg.DEDUP_BOXES = 0
 
-    cfg.TEST.BINARY = True
+    # Must turn this on because we use the test im_detect() method to harvest
+    # hard negatives
+    cfg.TEST.SVM = True
 
     args = parse_args()