Group-Wise Semantic Mining for Weakly Supervised Semantic Segmentation

Acquiring sufficient ground-truth supervision to train deep vi- sual models has been a bottleneck over the years due data-hungry nature of learning. This is exacerbated in some structured prediction tasks, such as semantic segmen- tation, which requires pixel-level annotations. work ad- dresses weakly supervised segmentation (WSSS), with goal bridging gap between image-level anno- tations and segmentation. We formulate WSSS novel group-wise learning task that explicitly se- mantic dependencies group images estimate more reliable pseudo ground-truths, can be used for training accurate models. In particular, we devise graph neural network (GNN) min- ing, wherein input are represented nodes, underlying relations pair char- acterized by an efficient co-attention mechanism. Moreover, order prevent model from paying excessive atten- tion common semantics only, further propose dropout layer, encouraging learn complete object responses. The whole end-to- end trainable iterative message passing, propagates interaction cues progressively improve performance. conduct experiments on popular PAS- CAL VOC 2012 COCO benchmarks, our yields state-of-the-art Our code available at: https://github.com/Lixy1997/Group-WSSS.