[Bug]illegal box Error

[SiyuanWei]

Prerequisite

• I have searched Issues and Discussions but cannot get the expected help.
• I have read the FAQ documentation but cannot get the expected help.
• The bug has not been fixed in the latest version (dev-1.x) or latest version (dev-1.0).

Task

I'm using the official example scripts/configs for the officially supported tasks/models/datasets.

Branch

main branch https://github.com/open-mmlab/mmdetection3d

Environment

After I processed the meta data of MMScan_QA.json, I noticed that there are a few illegal 3d boxes in it, which the size is below zero (if it is in [center,size ,zxy_angles] format). Are there any bugs in MMScanQA data generation?

Reproduces the problem - code sample

pass

Reproduces the problem - command or script

pass

Reproduces the problem - error message

pass

Additional information

No response

[SiyuanWei]

And I check the original MMScan_QA.json and found there are 174 items with illegal input_boxes in its training split

[rbler1234]

Special thanks for your feedback!
Since we hardly use input_bboxes but instead index the corresponding box from boxes given by MMScan dataloader tool based on input_bboxes_id, we did not encounter and overlook this issue. Actually, we recommend users to index the corresponding box from those given by MMScan dataloader tool based on input_bboxes_id (which is loaded from embodiedscan_infos_train/val.pkl, such that boxes can retain more precision, and we have strictly checked and verified). In the next release, we will remove the redundant key input_bboxes.

We checked all input_bboxes in the last release's QA and found that 55 out of over 100,000 objects had negative size values (e.g. object 91 from scene0000_00 is one of these 55 objects). This is because the size of the box when annotated is stored as a vector and the way to correct such issues is to take the absolute value (e.g. the correct bbox for object 91 is:

 [-2.11966847 -3.01424103  0.69299728  0.13252357  0.07769955 0.01025605 1.57079633  0.         -0.        ].

Although the proportion of such errors is extremely small, it may impact users who directly use input_bboxes in MMScan QA tasks. We have corrected the input_bboxes in MMScan_QA.json, aligned them with the boxes in embodiedscan_infos_train/val.pkl and will re-upload them to EmbodiedScan-v2-beta(the original link) under MMScan_samples/MMScan_QA.json these days.

Once again, thank you so much for helping us figure this out!