search

filaPro / oneformer3d

[CVPR2024] OneFormer3D: One Transformer for Unified Point Cloud Segmentation
Other
348 stars 32 forks source link

3d box detection? #44

Closed jiachen0212 closed 7 months ago

jiachen0212 commented 8 months ago

hello， thanks for your work~ However, I encountered some problems when using 3D point cloud detection~ I would like to ask, does your work support 3D detection and output box and orientation? I followed the scannet data format step by step, but encountered this bug ~ 1711341586091 lQDPJwiCFmwGrY12zQHhsHe8JRhVC8G5Be6YYfUh9gA_481_118 lQDPKHdus1Jjjo3MrM0Bg7ADQZi28amhfwXumGJ154UA_387_172

oneformer3d-contributor commented 8 months ago

At training stage this repo has nothing to do with object detection. However on test stage the object detection can be done simply with mapping each predicted instance with the smallest axis-aligned bounding box. The idea how to do it is described in #21.

jiachen0212 commented 8 months ago

At training stage this repo has nothing to do with object detection. However on test stage the object detection can be done simply with mapping each predicted instance with the smallest axis-aligned bounding box. The idea how to do it is described in #21.

Thank you very much for your reply. I modified 3 codes according to #21, with_bbox_3d=False during train and with_bbox_3d=True during test. But I still get this bug..

1711358366159

oneformer3d-contributor commented 8 months ago

Please import DepthInstance3DBoxes from mmdet3d.structures.