Abing: Adjusted Binarized Normed Gradients for objectness measure

Objectness measure, which generates some candidate object proposals, has been shown to accelerate the traditional sliding window category-dependent object detection methods. Binarized Normed Gradients (BING) is one of the state-of-the-art detectors. It achieves high object detection rate (DR), but moderate object overlap rate (OR) because the candidate proposals produced by BING are fixed-sized. In this paper, we propose a novel objectness detector, named as ABING (Adjusted Binarized Normed Gradients). It adjusts the fixed-sized proposals produced by BING, which can be accelerated by heap sort NMS (HS-NMS), and yields variable-sized ones by effectively exploiting Top Border None Object Boundary (TBNOB) principle and superpixel line integral (LI) cue. In experiments on the challenging PASCAL VOC 2007 dataset, we show that our ABING detector can consistently outperform BING with any number of proposals. Moreover, with well-chosen parameters, ABING can markedly enhance the DR and OR of BING with a certain number of proposals (e.g. 1000 proposals).