search

microsoft / CameraTraps

PyTorch Wildlife: a Collaborative Deep Learning Framework for Conservation.
https://cameratraps.readthedocs.io/en/latest/
MIT License
757 stars 239 forks source link

Implementing Oriented Bounding Box feature #507

Open acse-dc1823 opened 3 months ago

acse-dc1823 commented 3 months ago

Search before asking

Description

I would like to suggest an Oriented Bounding Box feature for the detector. Currently, all bounding boxes have parallel and perpendicular lines relative to the sides of the image. For an animal that is walking down, it would make the bounding box smaller and in the same orientation if this is implemented. https://docs.ultralytics.com/datasets/obb/

Use case

The best use case for this would be in animal re-identification Deep Learning models. They are very susceptible to pose changes, so having the animal in question always horizontal would be a good addition.

Additional

No response

Are you willing to submit a PR?

zhmiao commented 3 months ago

Hello @acse-dc1823, thanks for asking! This is a feature that will require specific annotations. If the data has this type of annotations, we can have visualization functions for such bounding boxes. However, the majority of the training data for MegaDetector are regular bboxes, therefore, it might be a little difficult to directly incorporate such feature to the current detection model. Maybe in the future, we will fine-tune a model with datasets of such type of bounding boxes.