CLIP is Also an Efficient Segmenter

表1

复现结果–Seed：70.7245673447014，dCRF：74.85437742935268
误差小于0.5个点，可以接受

表4

复现结果–训练300轮，Val：58.76741354153312，Test：59.18210

结论和配置

1、sharpness can serve as convenient guidance for prompt choice, and only image-level labels are needed.
2、Finally select “a clean origami {}.” as prompt，merge synonyms at the sentence level.
3、In this paper, leverage the ViT-based CLIP model.
4、A parameter λ is used to binarize the CAM，set λ to 0.4 for VOC.
5、Hyper-parameter Selection for µ in CGL：set µ to 0.95 in experiments.

说明

label[confidence < 0.95] = 255 #隐式执行CGL

感想

VOC全部复现完成，第一次啊！