Thanks to ToMe.ETM-RL made improvement based on ToMe.For details on ToMe, please visit http://github.com/facebookresearch/ToMe.
ToMe Pipeline: A straightforward BSM matching process that often leads to information loss, especially for foreground tokens.
ETM-RL Pipeline: Our method features a multi-stage feature enhancement pipeline, including adaptive contextual windows and fixed-radius neighborhood fusion, which significantly preserves critical information and structural integrity.
1、ETN-RL mainly reduces image information loss through two methods: an adaptive contextual window mechanism and a fixed-radius k-nearest neighbor strategy.We have provided the detailed code for these two methods in the merging.py file and applied them in the method called bipartite_soft_matching_xincheng to replace the original bipartite_soft_matching (BSM) method.If you only want to use one of the methods to enhance your image information, please comment out the others in the corresponding method.
2、We provide a test script -- test.py, which can compress the images you want to process. First, you need to modify the absolute path of the image you want to process, and then choose different r values to perform different degrees of merging. r=1 indicates a low degree of merging, while r=25 indicates a very high degree of merging.Then run：
python test.py