search

3dlg-hcvc / M3DRef-CLIP

[ICCV 2023] Multi3DRefer: Grounding Text Description to Multiple 3D Objects
https://3dlg-hcvc.github.io/multi3drefer/
MIT License
64 stars 3 forks source link

Train on ScanRefer dataset #19

Closed yihang-xdu closed 3 weeks ago

yihang-xdu commented 1 month ago

Dear Authors, Thanks for your great work! When I execute the command "python train.py data=scanrefer experiment_name=scanrefer +detector_path=checkpoints/PointGroup_ScanNet.ckpt Model.net work. The detector. Use_gt_proposal = True ", appear the following error:

File "/data1/zyh/M3DRef-CLIP/m3drefclip/data/dataset/scanrefer.py", line 35, in _load_language_data "eval_type": item["eval_type"], KeyError: 'eval_type'

File "M3DRef - CLIP/dataset/scanrefer/metadata/ScanRefer_filtered_train json" does not contain 'eval_type'.
I noticed that when training on the nr3d dataset, its 'eval_type' value is added as an extra in the code.https://github.com/3dlg-hcvc/M3DRef-CLIP/blob/babfcc4b4f06bb8ce1471513095a1c80acc0edb5/m3drefclip/data/dataset/nr3d.py#L41-L50. So I would like to ask whether it is necessary to add an extra 'eval_type' as when training on the nr3d data set?

eamonn-zh commented 3 weeks ago

Hi @yihang-xdu , I think you skipped step 2 for ScanRefer dataset preprocessing (please follow this instruction). The script will add the eval_type for you. This flag is only used for evaluation purposes.