Datasets:

ksikka
/

test

task_categories:
  - keypoint-detection
license: cc-by-4.0
tags:
  - biology
  - pose-estimation
  - multiview
  - fly
  - lightning-pose
pretty_name: Fly Anipose (Lightning Pose subset)
size_categories:
  - 1K<n<10K
configs:
  - config_name: default
    data_files:
      - split: ind
        path: data/ind-train-*.parquet
      - split: ood
        path: data/ood-test-*.parquet

Fly Anipose — Lightning Pose Multiview Dataset

6-camera pose estimation dataset for Drosophila leg keypoints, packaged for use with Lightning Pose.

Dataset Description

Head-fixed flies run on a spherical treadmill while 6 synchronized cameras capture locomotion at 300 Hz. Each frame is labeled with 30 keypoints — 5 joint segments (A–E) on each of 6 legs (left legs L1–L3, right legs R1–R3).

Labels are filtered Anipose predictions, not hand-labeled frames. They were constructed by:

Removing instances with mean 3D reprojection error > 10 px
Running k-means on 3D poses and keeping 25 instances per session
Using filtered 2D predictions; setting keypoints with 2D reprojection error > 10 px to NaN

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

Data Splits

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

InD and OOD sets contain different animals/sessions (no overlap).

CollectedData_Cam-{A-F}.csv — InD labels; videos/ — InD videos
CollectedData_Cam-{A-F}_new.csv — OOD labels; videos_new/ — OOD videos

Keypoints

30 keypoints total: side (L/R) + leg number (1–3) + segment (A–E, proximal→distal).

Left legs	Right legs
L1A, L1B, L1C, L1D, L1E	R1A, R1B, R1C, R1D, R1E
L2A, L2B, L2C, L2D, L2E	R2A, R2B, R2C, R2D, R2E
L3A, L3B, L3C, L3D, L3E	R3A, R3B, R3C, R3D, R3E

Directory Structure

fly_anipose_subset/
├── labeled-data/           # Extracted frames per session×view; includes ±2 context frames
├── videos/                 # Full InD session videos (<SessionKey>_<View>.mp4)
├── calibrations/           # Per-session camera calibration (.toml) for 3D features
├── calibrations.csv        # InD calibration index
├── calibrations_new.csv    # OOD calibration index
├── CollectedData_Cam-{A-F}.csv      # InD 2D keypoint labels (x,y per keypoint)
├── CollectedData_Cam-{A-F}_new.csv  # OOD 2D keypoint labels
├── config_fly-anipose.yaml # Sample Lightning Pose training config
├── project.yaml            # View and keypoint definitions (required by LP App)
└── models/                 # Pre-trained model checkpoints
    ├── baseline/
    ├── seed1/
    ├── seed2/
    └── pleasant_ensemble/

See the Lightning Pose documentation for full details on the multiview data directory structure and model 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: resnet50_animal_ap10k
Views: Cam-A through Cam-F
Keypoints: 30

Update data.data_dir and data.video_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 Anipose 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, Elisha and Azim, Eiman and
             Brunton, Bingni W and Tuthill, John C},
  journal = {Cell Reports},
  volume  = {36},
  number  = {13},
  year    = {2021},
  doi     = {10.1016/j.celrep.2021.109730}
}

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