search

facebookresearch / votenet

Deep Hough Voting for 3D Object Detection in Point Clouds
MIT License
1.7k stars 376 forks source link

Question about results on ScanNet. Why result of paper v2 is significantly higher than that of paper v1? #27

Open charlesCXK opened 5 years ago

charlesCXK commented 5 years ago

Hi, I have a question about the results on ScanNet dataset. I find that mAP@0.25 is reported to be 46.75% in the first version of your paper, while mAP@0.25 is boosted to be 58.6% in the version 2. Could you tell me how do you gain such an amazing increase? Thank you !

charlesCXK commented 5 years ago

We reproduced VoteNet according to the setting in the paper versio-1, and got a reasonable result on SUNRGBD. However, we only got 49.8@0.25 on ScanNet, which is reasonable when compared with paper version-1, but is far lower than the result in paper version-2. Thus we create this issue.

charlesq34 commented 5 years ago

Hi @charlesCXK

The major change is the learning rate scheduling. Originally we have exponential decay of the learning rate for paper version-1. Later we updated it to step-wise learning rate decay with the same intervals as that for SUN RGB-D training (80,120,160 epochs). The delayed decay of the learning contributed to the performance boost.

charlesCXK commented 5 years ago

Thanks! We will check it. @charlesq34