search

VISION-SJTU / USOT

[ICCV2021] Learning to Track Objects from Unlabeled Videos
63 stars 7 forks source link

Data set related issues #4

Closed zrz1018 closed 2 years ago

zrz1018 commented 2 years ago

The got10_flow.json file I downloaded directly from Google cloud drive only reached 9334 in the end, but the GOT10k training data set has 9335 files, may I ask what is the situation?

zhengjilai commented 2 years ago

In line 63 of parse_got10k_flow.py, you may notice that, if a failure of generating pseudo boxes (by flow and DP) happens, the preprocessing script will skip this video and never load this video into the json file.

In my experience, this exception is typically triggered if a runtime error, say "CUDA out of memory", occurs. I have once checked the generated pseudo box file, and found that only <10 videos have been skipped. This problem can be easily fixed with a dynamic processing batch for flow generation.

Frankly speaking, using slightly less videos (>9325) only has very subtle influence on the overall performance. Don't worry about that.