Missing config and hard-coded selection budget

[GerritBartels]

First off, thank you for your exciting contribution to the field of Active Learning for LiDAR-based 3D Object Detection!

While going through your code, I stumbled upon two things:

• The config for running STONE seems to be missing in the tools/cfgs/active-kitti_models folder. I assume it is identical to that of, e.g., the random config, yet I think to avoid confusion and/or mistakes, providing a config would be a good idea.

• The selection budget for each AL cycle seems to be hard-coded to 100 samples. Is there a specific reason for that or is it a residual from during development?

See pcdet/query_strategies/STONE_sampling.py line 375-386:

        # returned the index of acquired bounding boxes
        if len(selected_frames_final) < 100:
            # Calculate how many elements to extract from selected_frames
            elements_needed = 100 - len(selected_frames_final)

            # Extract the top elements_needed from selected_frames
            top_elements = selected_frames[:elements_needed]

            # Append extracted elements to selected_frames_final
            selected_frames_final.extend(top_elements)

        return selected_frames_final

Thank you in advance and kind regards, Gerrit

[RuiyuM]

Hi Gerrit Bartels,

Thank you for your feedback! I really appreciate you pointing out the issues with my configuration. I will definitely spend more time working on that aspect to improve it.

Regarding the budget hardcoded to 100, there isn’t a specific reason behind this choice—it simply follows the setup from the previous work (https://github.com/Luoyadan/CRB-active-3Ddet). However, you are absolutely free to adjust the budget. If your active learning method achieves good results with a lower budget, that would be a great improvement, and reducing the budget across all baselines would make sense.

Thanks again! Ruiyu