Cracking BING and Beyond
Qiyang Zhao, Zhibin Liu · 2014
The problem, generic objectness proposal, aims to reduce the candidate windows for object detection tasks. The popular evaluation criterion for related methods is detection-rate/windows-amount(DR-#WIN), where DR is the percentage of groundtruth objects covered by proposal windows. An object is considered “covered” by a window only if the strict PASCAL-overall criterion [3] is satisfied (the intersection of a proposal window and the object rectangle is not smaller than half of their union, so we call it “0.5-criterion” for short). Under the DR-#WIN evaluation framework, BING [2] in CVPR2014, obtains the best performance on the VOC2007 test set. It recalls 96.2% objects with only 1,000 proposal windows. The more surprising is the method is totally a realtime one. The authors of BING suggest that, after being resized to a fixed size (8× 8), almost all annotated rectangle regions share a common characteristics in gradients [2]. This commonness is captured by a template W learned from training images with a linear SVM. Besides this, the subtle differences between diverse width/height configurations are captured in a re-weighting model. Therefore BING consists of two stages: calculating W in stage I, and learning the re-weighting model in stage II. Furthermore, BING uses smart bitwise operations to calculate the inner product of W and candidate windows, so to improve the efficiency. We designed several templates by hands to substitute W , to verify whether templates play a key role in BING. These templates become less correlated to W in turn, but their performances on VOC 2007 test set are very close, see Fig.1.a. Next we discarded any templates and directly assigned the scores of stage I with uniformly random values (we call this method RAND-SCORE). Surprisingly, the performance of RANDSCORE is even very close to BING, as shown in Fig.1.b. It is clear that these templates do not have as strong significance as suggested in [2]. Then what on earth makes BING performing so well? To get the deep insight, we finished a theoretical analysis from the view of combinatorial geometry. We try to construct a small set of windows to “cover” all legal rectangles (we call it a full cover set). This is an atypical covering problem in combinatorial geometry [1]. We proposed four lemmas to solve it in the full paper. In conclusion, for an image of the width M and height N, we can use s(i, j)windows of the width 2i · √ 2 and height 2 j · √ 2 to cover all 2i ≤ w ≤ 2i+1,2 j ≤ h ≤ 2 j+1 rectangle regions, where s(i, j) = ⌈ M−2i ( 1− √ 2 2 ) ·2i· √ 2 ⌉ · ⌈ N−2 j ( 1− √ 2 2 ) ·2 j · √ 2 ⌉. Suppose the image size