Entire Deformable ConvNets for semantic segmentation

摘要

Deformable ConvNets achieve excellent performance due to learnable spatial sampling locations. However, the spatial support may deviate from the region of interest, according to observations of its behavior. More Deformable ConvNets reduce abnormal behavior mainly by relying on a modulation mechanism, but they do not identify the reason for the behavior. In this work, we find that the unusual behavior is caused by the interaction between offsets, and we present a novel method named Entire Deformable ConvNets for reducing its influence. This updated version helps the spatial support better focus on pertinent image content. In the new method, deformable convolution is replaced with a new module, namely, entire deformable convolution. The new module avoids mutual interference by changing the manner of offset allocation. Using this simple pattern, entire deformable convolution inherits all the advantages of deformable convolution and achieves better performance. Furthermore, this strategy is also applicable to More Deformable ConvNets. Similarly, Entire More Deformable ConvNets contain a new module called entire modulated deformable convolution. Extensive experiments on PASCAL VOC and MS COCO benchmarks show that our methods realize excellent improvement over the original versions.