Improved DSP matching with space-spanning scales and SLIC superpixels for dense scene correspondences

The Deformable Spatial Pyramid (DSP) Matching algorithm is popular for densely matching images with similar content but from different scenes. It achieves high matching accuracy by regularizing matching consistency at multiple spatial extends. Nevertheless, DSP computes dense SIFTs at a fixed scale across different spatial levels. Considering scale difference between images is crucial to matching accuracy, we propose to use space-spanning scales to improve DSP by extracting dense SIFTs at different scales for different pyramid level. Qualitative and quantitative experiments show our approach gains significant improvements, especially in dealing with images containing multiple objects from the same category. Moreover, SLIC superpixel algorithm is exploited to further refine the generated flow field. Experimental results show that this post-process is very efficient in reducing clutters and preserves much more texture information.