Regional Semantic Contrast and Aggregation for Weakly Supervised Semantic Segmentation
时间: 2024-02-06 10:03:21 浏览: 32
Weakly supervised semantic segmentation is a challenging task where only image-level labels are available for training. In this paper, we propose a novel approach that utilizes regional semantic contrast and aggregation to improve the performance of weakly supervised semantic segmentation.
First, we introduce a regional semantic contrast loss that encourages the model to focus on regions with high semantic contrast, i.e., regions where the difference between the foreground and background classes is high. Specifically, we compute the semantic contrast map by subtracting the predicted probability of the background class from the predicted probability of the foreground class, and then apply a Gaussian filter to smooth the map.
Second, we propose a regional aggregation module that aggregates the features of each region to obtain a more representative feature representation. The module consists of a region proposal network that generates candidate regions based on the semantic contrast map, and a region-based pooling layer that pools the features within each region.
Finally, we combine the regional semantic contrast loss and the regional aggregation module with a standard cross-entropy loss to train the model. Experimental results on the PASCAL VOC 2012 dataset demonstrate that our approach achieves state-of-the-art performance among weakly supervised methods, and even outperforms some fully supervised methods.