Got low accuracy results for Scannet per-point prediction

[sylyt62]

I followed ScanNet Segmentation instructions, but got bad prediction results.

Firstly, I downloaded scene0000_00 from training set and scene0707_00 from test set, ran python tools/scannet.py --run process_scannet --path_in <scannet_folder> to prepare the data, and ceated my own filelist.txt

Secondly, changed config file (based on seg_scannet_eval.yaml, only changed in and out dir) and ran python segmentation.py --config configs/seg_custom_eval.yaml, using weights 00600.model.pth

Finally, converted the predictions to segmentation labels.

No errors jumped out, and got results in logs folder. But after I attach the predicted labels to ply file, seems like it can only distinguish walls and floor.

Did I miss something?

pic1: scene0000_00 predictions

pic2: scene0000_00 labels

scene0707_00 predictions

[Peng-Shuai Wang]

Could you forward the network on the validation dataset and get a mIoU around 74.0?