xinw1012 / cycle-confusion

Code and models for ICCV2021 paper "Robust Object Detection via Instance-Level Temporal Cycle Confusion".
https://xinw.ai/cyc-conf
Apache License 2.0
74 stars 9 forks source link

Can not reach the reported performance #4

Open fanq15 opened 2 years ago

fanq15 commented 2 years ago

Hi, thanks for this excellent work! I try to train the model using the config: R50_FPN_daytime_cycle_conf.yaml on BDD100k. However, the performance is only 18.126 AP which is the same as my retrained baseline (18.108 AP). There is a 1.0 AP gap with the reported performance in the paper.

Linwei-Chen commented 1 year ago

I have a similar problem, moreover, the baseline of R50_FPN_daytime.yaml can achieve 18.2 AP while R50_FPN_daytime_cycle_conf.yaml ahcievs 18.7 AP.