search

Robertwyq / LSMOL

[NeurIPS 2022] 4D Unsupervised Object Discovery
52 stars 0 forks source link

Evaluation 3D instance segmentation #4

Open chreisinger opened 10 months ago

chreisinger commented 10 months ago

Hello,

thank you for your excellent work! I have two questions about your 3D instance segmentation:

Please give me your thoughts on both points.

Thank you in advance!

Robertwyq commented 10 months ago

Thanks for your attention. We will carefully examine the issues you have raised and provide you with a response.

chreisinger commented 6 months ago

Hello, are there any updates regarding this issue?

I have another question regarding the case of true positives: https://github.com/Robertwyq/LSMOL/blob/d35cb4a58764ef57b6f47082101c643bf3bd5013/evaluation/eval_3d_instance.py#L121-L129

There is no handling for the case of two clusters assigned to the same ground truth instance (with valid IoU, e.g., 0.4). In that case, both clusters would count as TPs, and the FN count would be negative. https://github.com/Robertwyq/LSMOL/blob/9a39b26804285d9a03e56664a34abab187c38ff4/evaluation/eval_3d_instance.py#L130

Do I miss here something?

Thank you for your time.

Best regards, Christian