Thie repo provides the official implementation of our AAAI-2023 paper “SSDA3D: Semi-supervised Domain Adaptation for 3D Object Detection from Point Cloud”.
In the paper there are comparisons with ST3D in both UDA and SSDA settings. I have a few questions about these experiments.
Since this repository does not seem to contain code related to ST3D, did you modify the official ST3D release to use CenterPoint?
For waymo->nuscenes experiments, the ST3D configuration has pseudo-label updates disabled (UPDATE_PSEUDO_LABEL_INTERVAL=1000), were you using the same settings as them?
Hello, thanks for your great work.
I found some differences between the code and paper. According to the paper, "before merging the 3D boxes, collision detection will be performed". But in the intra_domain_point_mixup function, the "gt_boxes" were merged directly. This may cause some problems.
Thanks for sharing this excellent work! Could you please share the details about how the 2nd stage works under 100% labeled target setting? In this case, there will be no TU data, will the student-teacher learning still be conducted? Thanks in advance!
How did you generate the partial_info_data (1%, 5%, 10%, 20%)? For example, for the 1% data, did you take the first 1% of the info from the complete info.pkl file and split it into the top 1% and the last 99% of the pkl file? Or did you randomly sample 1% of the total amount of samples?