A Hybrid Approach for Object Proposal Generation

Abstract

Object detection in natural images is evolving, with enormous commercial achievements, becoming relatively common in every industry. Modern research in this area is progressing in many directions, with numerous different techniques being proposed to achieve state-of-the-art detection performance. Recent object detection methods use two steps to detect high-quality objects: first, it generates a set of object proposals as accurate as possible, and then these proposals are passed to object classifier for post-classification. This paper presents an efficient new hybrid object proposal method, which gets the initial proposal by computing multiple hierarchical segmentations using super pixels and then ranks the proposal according to region score – which is defined as number of contours wholly enclosed in the proposed region, passing only the top object proposal for the post-classification. Passing few object proposals in the object detection pipeline for post-classification speeds up the object detection process. This paper demonstrates that our method results in high-quality class-independent object locations, with mean average best overlap of 0.833 at 1500 locations, resulting in a superior detection rate in object detection tasks at relatively fast speeds – as compared to object detection methods using selective search – and greatly reduces the false-positive rate.