FSS
FSS copied to clipboard
COCO-MS setting encountered issues
Among the four tasks (VOC-SS, VOC-MS, COCO-SS, COCO-MS), only the COCO-MS setting had a significant difference from the indicators in the paper. We adopted lr=0.001 and iter=100 (consistent with the Supplementary file). Under the 1 shot setting, we only obtained miou-b=39 and miou-n=7.9 (40.4 and 10.4 in the original paper). We have been debugging for a month now. Could you please give us some suggestions?
This is our coco-ms.sh file: