Official code for Spyropose: SE(3) pyramids for object pose distribution estimation, ICCVW 2023.
Project page.

We recommend using uv for dependency management. In a project, add spyropose as a dependency, for example from a git repo as below

uv add REPO_URL

Spyropose trains a model per object.

Store your training data in the BOP format in ./data/bop/DATASET

uv run -m spyropose.scripts.inspect_data \
  --data.obj.dataset=DATASET \
  --data.obj.obj=OBJECT_ID_OR_NAME

If the data looks fine, you can train a spyropose model with the below command. This would use scenes with index 0 through 18 for training and scene 19 for validation.

uv run -m spyropose.scripts.train \
  --obj.dataset=DATASET \
  --obj.obj=OBJECT_ID_OR_NAME \
  --data_train.scene_id_range="[0,19]" \
  --data_valid.scene_id_range="[19,20]"

A script to train a simple detector is included and can be trained with a similar script:

uv run -m spyropose.detection.train \
  --obj.dataset=DATASET \
  --obj.obj=OBJECT_ID_OR_NAME \
  --data_train.scene_id_range="[0,19]" \
  --data_valid.scene_id_range="[19,20]"

See ./src/spyropose/scripts/infer.py.

Can be run like so:

uv run -m spyropose.scripts.infer \
  ./data/spyropose_detector/2kewoepx ./data/spyropose/dwq4lb0a 19 0

Check out initial commit.

@inproceedings{haugaard2023spyropose,
  title={Spyropose: Se (3) pyramids for object pose distribution estimation},
  author={Haugaard, Rasmus Laurvig and Hagelskj{\ae}r, Frederik and Iversen, Thorbj{\o}rn Mosekj{\ae}r},
  booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision},
  pages={2082--2091},
  year={2023}
}