search

hustvl / CrossVIS

[ICCV 2021] Crossover Learning for Fast Online Video Instance Segmentation
https://arxiv.org/abs/2104.05970
Other
85 stars 4 forks source link

x3 Schedule Results #7

Closed JialianW closed 2 years ago

JialianW commented 2 years ago

Dear author,

Thanks for your great open-source work. In the literature that image object detectors on COCO usually try an x3 training schedule. I wonder if you ever trained the method for an x3 training schedule for VIS? Will it further improve results or already saturate in the x1 schedule?

Yuxin-CV commented 2 years ago

Hi@JialianW

We find longer schedule results in even worse results.

Actually, for VIS dataset, usually 5-6 epochs (i.e., 0.5x schedule) finetuning works the best.