junjiehe96 / FastInst

[CVPR2023] FastInst: A Simple Query-Based Model for Real-Time Instance Segmentation
MIT License
178 stars 16 forks source link

Not satisfied with overall results of the model. #36

Open sparshgarg23 opened 7 months ago

sparshgarg23 commented 7 months ago

So I was able to finish the training of the model for roughly 280,000 iterations and I am not satisified with the overall results of the model. The overall mean AP for resnet50 with batch size set to 6 comes out to be only 27% with the AP for the smaller objects coming out to be only 9%.

Any reasons why this drop in AP is happening,is it because of the batch size . I noticed that whenver I resume training,instead of resuming from where it left off,it starts at iteration 0. An example of the instance segmentation on coco image is shown below Original image 000000071074 Result image img1