Skip to content

Problems encountered in improving the pointpillar model #3129

@Allcaros

Description

@Allcaros

What is the feature?

I have conducted the default pointpillar training according to the official documentation. Although there are some differences in the results, they are acceptable.
The problem I encountered was that when I intended to use other model components in mmdet3d, problems of varying degrees occurred. It took me a long time to get the code running, but the result was disastrous. My idea is to use the centerpoint model and the centerpoint_head detection head to replace the original components of pointpillar;
Now when I start thinking from the beginning, I feel that there seem to be some problems in many aspects that I have overlooked. If you can provide some ideas, I will:

  1. I realized that after generating the.pkl file using create_data.py, a kitti_gt_database would be generated. This seems to be a folder for increasing the dataset, but I haven't used it. Maybe do you have any ideas for me to use this folder correctly?

  2. In the default model code, there is a parameter called use_ground_plane, which is set to True by default. However, when I run it, it prompts me that there is no related file. So, I simply set it to flase. When I started thinking from the beginning, I realized that this file should be used to constrain the ground target To enable different category targets to be better positioned for learning, if I want to avoid the above-mentioned error reports, where should I download the relevant use_ground_plane file, and what should the training content or dataset placement look like?

  3. The results after I replaced the detection head are as follows. I want to know why the ap of the car category is so low when there are so many of them and they are so large in volume. What I can think of is that kitti does not provide a center point coordinate annotation, so such large targets cannot return to the center point and thus cannot be detected. Maybe you have a better idea. Please tell me.

Image

4、Theoretically speaking, it is feasible for me to replace it with components of the centerpoint network. But why is the result so bad? Maybe I need better code optimization, but currently I doubt the compatibility issue. Is incompatibility worse than feasibility?

If you have better ideas or other answers, please let me know. Thank you very much

Any other context?

No response

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions