Results on ScanNet

[raoyongming]

Hi, thanks for sharing your code.

I am trying to reproduce the results on ScanNet using the default configs. But I can only obtain around 64-65 mAP@0.25, which is lower than the results in your paper (67.2). Do you have any advice to reproduce your results?

Here are the training logs log_train.txt log_train_run2.txt

[zaiweizhang]

I briefly looked at your log files. I do think that the performance is different.
Here are two log files from our experiments.
log_train.txt
log_train_2.txt

Would you mind sharing your pytorch, conda, and cudnn version?

[raoyongming]

Thanks for your quick reply. My experiments were conducted using 4 1080ti gpus with pytorch 1.6.0, cuda 10.2 and cudnn 7.6.5

[zaiweizhang]

Since we are using Votenet's codebase, most of our experiments are conducted with pytorch 1.1.0. cuda 10.0, and cudnn 7.6
I have also tried with pytorch version 1.2.0, and cuda 10.1. The results are similar.

I would suggest you try our code with this setting: pytorch 1.1.0. cuda 10.0, and cudnn 7.6.

[raoyongming]

I tried using pytorch 1.1.0 and 3 gpus (bs=9) to conduct the experiments and obtained 67-68 mAP@0.25 (slightly better than the results in your paper). Maybe it is because batch size = 2 on each gpu is too small for BN. Thanks for your help!