search

happinesslz / SEED

[ECCV 2024] A Simple and Effective 3D DETR in Point Clouds
54 stars 0 forks source link

Binary classification for quality query selection? #2

Closed zhanggang001 closed 1 month ago

zhanggang001 commented 1 month ago

Dear authors,

I noticed the line tgt["labels"].fill_(0) for tgt in bin_targets in seed_head.py. Does this indicate that a binary classification is conducted for quality query selection?

Thank you in advance!

happinesslz commented 1 month ago

@zhanggang001 Great question! I carefully check our code, and find the line of ' tgt["labels"].fill_(0) for tgt in bin_targets' should be removed. As for quality query selection, we adopt a multi-class classifier rather than a binary classifier. We'll fix this issue and rerun the experiment. I anticipate the results will improve once this bug is addressed. Thanks again!

happinesslz commented 1 month ago

@zhanggang001 Hi, we re-run this experiment (6 epoch with 100% training data for the small model) after solving this bug and find there brings about 0.2-0.3 mAPH/L2 improvement!

zhanggang001 commented 1 month ago

@zhanggang001 Hi, we re-run this experiment (6 epoch with 100% training data for the small model) after solving this bug and find there brings about 0.2-0.3 mAPH/L2 improvement!

Thanks for your dedicated effort!!