search

facebookresearch / unbiased-teacher

PyTorch code for ICLR 2021 paper Unbiased Teacher for Semi-Supervised Object Detection
https://arxiv.org/abs/2102.09480
MIT License
409 stars 84 forks source link

Uploaded PASCAL VOC weight's cocoEVAL performance #74

Closed honggyuchoi closed 2 years ago

honggyuchoi commented 2 years ago

Hello @ycliu93, Thank you for sharing awesome work!

I have question about the performance of PASCAL VOC(uploaded model weight that using voc07 as labeled data and voc12 as unlabeled data)

When i check the performance(AP, AP50) of uploaded PASCAL VOC model weight with cocoEVAL, the AP is lower than the result that reported in the paper(paper -> AP: 48.7, AP50: 77.4 / uploaded model weight -> AP: 45.2, AP50:72.2)

image

Is it because of the small batch size? or is there any other issue?

Since some works(i.e. Instant Teaching) achieve similar performance(AP: 48.7, AP50:78.3) using only EMA and strong augmentation with small batch size(label:8, unlabel:8), i think that there could be other issue.

image

Last row is reimplementation result based on your repository. Reimplementation result also shows similar performance.

Thank you!

honggyuchoi commented 2 years ago

I made the mistake of evaluating the performance of the student model. image

Above image shows teacher's performance that evaluated on cocoEVAL.