Mining Better Samples and Semantic Consistency for Contrast Learning in Forest Semantic Segmentation

Image segmentation has impressive progress in the past several years. But good segmentation usually follows pixel-wise well-annotated labels which is energy-consuming. Moreover, the robustness would not be guaranteed due to lack-of-diversity datasets. Previous work usually focuses on pixels individually and pay less attention to the neighbor pixels. The local context would be scarce and the global context is not utilized following these ways. We proposal a method, named Forest Semantic Segmentation Network(FSSNet) to address these issues. FSSNet organizes original version and augmented version of images, as two inputs into student branch and teacher branch, and force the two outputs being consistent to strengthen the robustness of our model. Moreover, we not only consider pixel itself and also the neighbor pixels because the context of neighbor pixels helps understanding the pixel. FSSNet utilizes contrastive loss with memory bank to involve global context in training which will make pixels closer to others in same category and far away from pixels of different categories. A bank filter is suggested to improve the quality of features in the memory bank. we also proposal a new sample strategy to improve the effect of contrastive loss and reduce the computation. Our method can improve accuracy and strengthen the robustness with affordable extra computation during training process, and no additional computation during inference toward baseline. Comparing to benchmark, the proposed approach can improve the mIoU by 3.1% on our challenging dataset.