A Stochastic Approach for Selective Search Algorithms

This paper introduces a new algorithm that generates candidate proposals for an object detection pipeline. We introduce Stochastic Selective Search (SSS), a segmentation based selective search method, which differs from previous work in two ways. First and most importantly, SSS is much faster than current state-of-the-art algorithms while maintaining comparable accuracy. This is a result of our efficient stochastic segment merging process. Other work requires the computation of features to determine the order in which segments are merged. We show that currently used features from other work does not improve the results of SSS significantly and are therefore omitted. This makes our algorithm nearly twice as fast as the fastest prior selective search algorithms. Secondly, due to the stochastic merging process of SSS, it is not critically affected when two wrong segments are merged during the merging process, which leads to object proposals of higher quality. We show that SSS outperforms existing deterministic selective search methods while generating the same amount of proposals in less time. Additionally, we demonstrate the performance of our SSS algorithm in a state-of-the-art object detection pipeline based on convolutional networks.