Object Proposals Based on Variance Measure

Object proposals have recently become an important part of the object recognition process. Current object proposals are mostly based on hierarchical grouping which generates too many proposals and consumes too much processing time. This paper presents a framework utilizing variance measure to produce object proposals. We divide complete image into small patches. Our algorithm identifies possible object patches and merged them together to form object proposals. We evaluate our algorithm on UT interaction dataset. Experimental results show that our method generates fewer but quality proposals. Our method also performs reasonably fast than the state-of-the-art approaches.