performance degrade when applying a consistent evaluating code with other methods (TCL, GroupVit..)

[wuwuwuxxx]

Have you tested the performance of the proposed method with a consistent evaluating code of other methods (such as TCL, GroupViT)? After doing so, I observed a significant decrease in performance in the test results. For example, VOC20: 77.0, ADE20K: 14.3.

[wangf3014]

I'm not quite sure about how you conducted your evaluation. Can you provide more details about that? Did you import the SCLIP model into TCL's repo and test? Some mis-implementations would highly affect the results.

[wuwuwuxxx]

I incorporated the modifications from SCLIP's final layer into CLIP and followed TCL's testing code(main:validate_seg), including dataset configuration, testing procedures, and evaluation code (dataset.evaluate). I observed performance differences due to variations in image testing sizes and templates. However, even after aligning the image sizes and templates, the test results still differed from those reported in the paper, such as achieving 21.1 for cocostuff.