LAVT: Language-Aware Vision Transformer for Referring Image Segmentation

Referring image segmentation is a fundamental vision-language task that aimsto segment out an object referred to by a natural language expression from animage. One of the key challenges behind this task is leveraging the referringexpression for highlighting relevant positions in the image. A paradigm fortackling this problem is to leverage a powerful vision-language ("cross-modal")decoder to fuse features independently extracted from a vision encoder and alanguage encoder. Recent methods have made remarkable advancements in thisparadigm by exploiting Transformers as cross-modal decoders, concurrent to theTransformer's overwhelming success in many other vision-language tasks.Adopting a different approach in this work, we show that significantly bettercross-modal alignments can be achieved through the early fusion of linguisticand visual features in intermediate layers of a vision Transformer encodernetwork. By conducting cross-modal feature fusion in the visual featureencoding stage, we can leverage the well-proven correlation modeling power of aTransformer encoder for excavating helpful multi-modal context. This way,accurate segmentation results are readily harvested with a light-weight maskpredictor. Without bells and whistles, our method surpasses the previousstate-of-the-art methods on RefCOCO, RefCOCO+, and G-Ref by large margins.