How to evaluate the mAP(fast) in VID task?

[hahapt]

For VID methods like FGFA and SELSA, the author show the results of mAP(slow), mAP(medium) and mAP(fast) in the paper. When testing with mmtracking, I found here are results of AP(small), AP(medium) and AP(large), so how can we know the model works on object with different speed?

[JingweiZhang12]

We don't support this feature to analysis the performance of detecting objects with different motion speed yet, but you can implement it on our codebase. Good Luck!