【计算机视觉】Object Proposal之BING++

【计算机视觉】Object Proposal之BING++

本文是对 BING 算法的升级，主要是在快的同时保持定位精度 
两个 + 分别对应： edge-based recursive boxes as one “+”, and MTSE-based superpixel merging as the other “+”

Object Detection Recall (DR)：which is the ratio of the number of correctly detected objects and the total number of objects in the dataset

Proposal Localization Quality measured in terms of average best overlap (ABO)

mean average best overlap (MABO) 

Key insight: 我们注意到好的候选区域框是刚刚包含物体，edge 特征可以用于检测物体边缘

2 Problems in BING 
poor proposal localization quality

3 Our Solution: BING++ 
(1) We take proposals from BING as input. 
(2) We then recursively update current bounding boxes based on their current locations and surrounding edge points to form new bounding boxes. This process is recursed until we find 
(3) We finally apply fast super-pixel merging techniques to further refine the output bounding boxes of (2) and output final object proposals

算法流程： 

结果对比： 

转载于:.html