Regional Semantic Contrast and Aggregation for Weakly Supervised Semantic Segmentation
时间: 2024-02-06 22:03:32 浏览: 25
Weakly supervised semantic segmentation is a challenging task where only image-level labels are provided instead of pixel-level annotations. One approach to tackle this problem is to use regional semantic contrast and aggregation.
Regional semantic contrast involves identifying regions in the image that have a high contrast in semantic labels. This is done by computing the difference between the maximum and minimum probability of each class within a region. Regions with high contrast are likely to contain the object of interest.
Aggregation involves combining multiple weak segmentation maps generated from different regions of the image. This is done by assigning a weight to each map based on its regional semantic contrast. Maps with higher contrast are given higher weights, indicating that they are more likely to contain the object of interest.
The final segmentation map is obtained by combining the weighted maps using a simple weighted average. This approach has been shown to achieve state-of-the-art performance on several weakly supervised semantic segmentation benchmarks.