VNext
VNext copied to clipboard
Difference between ECCV paper and YouTube-VIS challenge
Hello, thank you for release your great work. I want to know training difference between ECCV paper and YouTube-VIS challenge. The mAP_L in your YouTube-VIS challenge paper is much higher than the checkpoint in this repo (50.7 vs 44.6). I try to use training settings in YouTube-VIS challenge paper to train IDOL (iter:12000, step:6000, lr:0.0002) while the result is even worse than the checkpoint in this repo. So I wonder why the mAP_L in your YouTube-VIS challenge paper is so high? Looking forward to your reply, thank you very much! Best.
Except iter, step and lr, I use default settings for other. I also use 8*A100 to train.