Skip to content

mAP Metric for 3D Bounding Boxes #3398

Description

@lhahner

🚀 Feature

I would like to be able to compute mean-average-precision for 3D Bounding boxes to also be able to evaluate detection systems that are capable of processing 3D data, like point-clouds, voxels or other.

Motivation

Currently whenever I want to compute the mAP with torchmetrics.detection.mean_ap.MeanAveragePrecision I am restricted to 2D data, meaning the parameter num_boxes is restricted to shape (num_boxes, 4).

Pitch

I would recommend to add a separate mAP class that only covers 3D bounding boxes. So to separate between
3D mAP and 2D mAP computation. The reason is that in 3D we have different algorithms to compute IoU and average precision etc.

Alternatives

  • Either extend the already existent torchmetrics.detection.mean_ap.MeanAveragePrecision.
  • Or implement a complete new metric class.

Additional context

Read more about mAP here:

Metadata

Metadata

Assignees

No one assigned

    Labels

    enhancementNew feature or request

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions