AUC on OTB100 0.54， training on VID and ytbb

STVIR / pysot

SenseTime Research platform for single object tracking, implementing algorithms like SiamRPN and SiamMask.

Apache License 2.0

4.44k stars 1.11k forks source link

AUC on OTB100 0.54， training on VID and ytbb #96

Closed xi-mao closed 5 years ago

xi-mao commented 5 years ago

I trained siamrpn_alex_dwxcorr_16gpu on VID and ytbb. but it's only got AUC 0.54 on OTB100. I train twice，and the results is same。I didn't change the hyperparameters。 I think the reason is that the hyperparameters is set for traning four datasets。Maye it’s not suitable for two datasets。 Anyone can explain to me？

ghost commented 5 years ago

You can change some hyperparameters like epoch in config.yaml and the VIDEOS_PER_EPOCH in pysot.core.config.py, matching with your datasets.

charles-str commented 5 years ago

I trained siamrpn_alex_dwxcorr_16gpu on GOT-10k train dataset, it's performance better than yours, but it can't catch up with author's ,can we communicate with each other?

liuyun19951104 commented 5 years ago

can you present how the VID dataset is? i cannot download it, and i want to train in my own video dataset. i don't know how to do?