paninski-lab/fly-anipose

Fly Anipose — Lightning Pose Multi-View Dataset

Multi-camera pose estimation dataset for fly leg keypoints, packaged for use with Lightning Pose.

Dataset Description

Head-fixed flies behave spontaneously on an air-supported ball, captured by six cameras at 300 Hz (Karashchuk et al., Cell Reports 2021). Frame sizes vary by view but are approximately 450 × 525 pixels, and are resized to 256 × 256 during training. Each frame is labeled with 30 keypoints: 5 joints on each of 6 legs.

Note: Labels in this dataset are not hand-labeled. They are filtered Anipose 3D predictions reprojected to each camera view (see Label Construction below).

Source data: Karashchuk et al., Cell Reports 2021 — original archive at https://doi.org/10.5061/dryad.nzs7h44s4.

Label Construction

For a subset of sessions in the Anipose data repository that contain 3D pose predictions, labels were constructed as follows:

1. Remove any instance where average 3D reprojection error is > 10 pixels.
2. Run k-means clustering on the remaining 3D poses (25 clusters per session) and select one instance per cluster.
3. Use the filtered 2D reprojections (not the original predictions) as labels.
4. Set any keypoint where per-keypoint 2D reprojection error > 10 pixels to NaN.

Data Splits

Split	Labeled instances	Sessions
In-distribution (InD)	377	16
Out-of-distribution (OOD)	300	12

InD and OOD sets contain different sessions / animals (no overlap). An "instance" is the set of synchronized frames across all six cameras at a single point in time.

• CollectedData_{cam}.csv — InD labels; videos/ — InD videos
• CollectedData_{cam}_test.csv — OOD labels; videos_test/ — OOD videos

Keypoints

30 keypoints per frame: 5 joints (A–E, proximal to distal) on each of 6 legs.

Left legs	Right legs
L1A–L1E (foreleg)	R1A–R1E (foreleg)
L2A–L2E (midleg)	R2A–R2E (midleg)
L3A–L3E (hindleg)	R3A–R3E (hindleg)

Directory Structure

fly-anipose/
├── calibrations/                       # Per-session camera calibration files (.toml)
├── labeled-data/                       # Extracted frames per session; includes ±2 context frames
├── videos/                             # InD session video clips
├── videos_test/                        # OOD session video clips
├── videos-for-each-labeled-frame/      # 51-frame videos centered on each OOD labeled frame
├── CollectedData_{cam}.csv             # InD 2D keypoint labels (x,y per keypoint) for each camera view
├── CollectedData_{cam}_test.csv        # OOD 2D keypoint labels for each camera view
├── config_fly-anipose.yaml             # Sample Lightning Pose training config
└── project.yaml                        # View and keypoint definitions (required by LP App)

The videos-for-each-labeled-frame/ directory contains 51-frame video clips with the labeled frame at the center, intended for use with temporal smoothers such as the Ensemble Kalman Smoother.

See the Lightning Pose documentation for full details on the multi-view data directory structure.

Usage with Lightning Pose

The included config_fly-anipose.yaml is a ready-to-use training config. Key settings:

• Image resize: 256 × 256
• Backbone: vits_dino
• Keypoints: 30
• Views: Cam-A, Cam-B, Cam-C, Cam-D, Cam-E, Cam-F

Update data.data_dir to absolute paths on your machine before training.

litpose train config_fly-anipose.yaml

Citation

If you use this dataset, please cite the original paper:

@article{karashchuk2021anipose,
  title={Anipose: A toolkit for robust markerless 3D pose estimation},
  author={Karashchuk, Pierre and Rupp, Katie L and Dickinson, Evyn S and Walling-Bell, Sarah and Sanders, Elischa and Azim, Eiman and Brunton, Bingni W and Tuthill, John C},
  journal={Cell reports},
  volume={36},
  number={13},
  year={2021},
  publisher={Elsevier}
}

Original data archive: https://doi.org/10.5061/dryad.nzs7h44s4.