search

facebookresearch / ContrastiveSceneContexts

Code for CVPR 2021 oral paper "Exploring Data-Efficient 3D Scene Understanding with Contrastive Scene Contexts"
MIT License
217 stars 29 forks source link

Scannet Semantic Segmentation Results #44

Open YilmazKadir opened 1 year ago

YilmazKadir commented 1 year ago

Hi, I have downloaded the pretrained model from model zoo from this link: http://kaldir.vc.in.tum.de/3dsis/contrastive_scene_contexts/pretrain/partition8_4096_100k.pth and trained the model for semantic segmentation on Scannet initialized with pretrained weights. I got 72.7 mIoU. Is this within the standard deviation? I used 8 GPUs with 48 batch size for 20k iterations. I did not change anything in the config file. Here is my log file: ddp_main.log

Sekunde commented 1 year ago

There could be variance, but 72.7 is a bit low. Are you using ME0.4.3?

YilmazKadir commented 1 year ago

Yes.

Sekunde commented 1 year ago

I re-run the scannet semantic segmentation with the released partition8_4096_100k.pth pre-trained weights, and got 73.4 which should be within the deviation, see the log: ddp_main.log I never saw the number was below 73 actually, but it does not mean 72.7 is impossible. The difference I can think of now could be the data used, let me upload the data we generated. Do you have the same version pytorch/cuda etc? not sure that could be a reason. Btw the released partition8_4096_100k.pth was a re-trained model after code refactoring, since you raised this up, I will also test the pre-trained models before refactoring.

YilmazKadir commented 1 year ago

I use the same pytorch and cuda. It would be great if you could upload the data you generated.

Sekunde commented 1 year ago

check this out: https://drive.google.com/file/d/1UtVXBPOiLnczQCdM2-0WouS3zRL8ldzw/view?usp=sharing

Sekunde commented 1 year ago

I re-run the scannet semantic segmentation using the weights pre-trained before code refactoring and got 73.8 (at iteration 19500), see the log: ddp_main.log