Semi-Global Context Network for Semantic Correspondence

摘要

Estimating semantic correspondence between pairs of images can be challenging as a result of intra-class variation, background clutter, and repetitive patterns. This paper proposes a convolutional neural network (CNN) that attempts to learn rich semantic representations that contain the global semantic context to enable robust semantic correspondence estimation against intra-class variation and repetitive patterns. We introduce a global context fused feature representation that efficiently employs the global semantic context in estimating semantic correspondence as well as a semi-global self-similarity feature to reduce background clutter-induced distraction in capturing the global semantic context. The proposed network is trained in an end-to-end manner using a weakly supervised loss, which requires a weak level of supervision involving annotation on image pairs. This weakly supervised loss is supplemented with a historical averaging loss to effectively train the network. Our approach decreases running time by a factor of more than four and reduces the training memory requirement by a factor of three and produces competitive or superior results relative to previous approaches on the PF-PASCAL, PF-WILLOW, and TSS benchmarks.