Very impressive work. I currently have some questions I need to ask you. I am currently using the open-source weights on your HuggingFace and modified the runs/run.sh work-dir to the corresponding governance folder. Then I changed the s1 and s2 model paths of the config files in the running script to None according to the comments, and reproduced the evaluation results. I used a single A800 card running the Segval evaluation. However, I found that many of my benchmark indicators and thesis metrics did not quite match, and I didn't know what the problem was
The above are some metrics for RefCOCO and ReasonSeg. The RefCO series dropped by an average of 1.5%, while ReasonSeg did not match.
Very impressive work. I currently have some questions I need to ask you. I am currently using the open-source weights on your HuggingFace and modified the runs/run.sh work-dir to the corresponding governance folder. Then I changed the s1 and s2 model paths of the config files in the running script to None according to the comments, and reproduced the evaluation results. I used a single A800 card running the Segval evaluation. However, I found that many of my benchmark indicators and thesis metrics did not quite match, and I didn't know what the problem was