Open3DIS
Open3DIS copied to clipboard
code for the ScanNet++ dataset
hi guy, thanks for great work here! I am try to do the experiments that you mentioned in the supplyment materials, but i came to notice that your code does not include the ScanNet++ dataset. BOTH in the ISBNet extraction code and dataset preparation part!! For example: THERE IS NOT segmenter3d/ISBNet/dataset/scannetpp dir for me to pre-process the dataset as a result , I have no idea how to generate the corresponding groundtruth file, i.e. "*_inst_nostuff.pth"
will you release that part of code soon? Many thanks again!
Hi @LiangXu123, Training the 3DIS network for ScanNet++ is a bit challenging for ISBNet as we are investigating proper hyperparams best suited for ScanNet++. ISBNet pretrained on ScanNet200 yields worse results on ScanNet++ so we did not include the inference code. as we are planning to release ISBNet training code (configs) for ScanNet++ in the future. Make sure you keep an eye out! Currently, we just use the 2D-only version for ScanNet++ and it achieves top-4 on the 3DIS benchmark.
Best, PhucNDA.
Hi @LiangXu123, Training the 3DIS network for ScanNet++ is a bit challenging for ISBNet as we are investigating proper hyperparams best suited for ScanNet++. ISBNet pretrained on ScanNet200 yields worse results on ScanNet++ so we did not include the inference code. as we are planning to release ISBNet training code (configs) for ScanNet++ in the future. Make sure you keep an eye out! Currently, we just use the 2D-only version for ScanNet++ and it achieves top-4 on the 3DIS benchmark.
Best, PhucNDA.
THANKS for your reply, But I stt you released the/configs/scannetpp.yaml
and I want to evaluate on the scannet++ dataset, could you please tell me how to generate the gr file for scannet++ dataset?:
gt_pth: './data/Scannetpp/Scannetpp_3D/val/groundtruth' # gt path
as it is required for the evaluation code?
OR may you you can just upload the groundtruth.tar ?