{"success":true,"data":{"author":"Tuthill Lab, University of Washington (source data); Paninski Lab, Columbia University (packaging)","bucketAccessType":"public","bucketName":"paninski-lab/fly-anipose","bucketPrefix":"","category":"Neuroscience","cloudProvider":"huggingface","createdAt":"2026-11-21T20:45:58.819Z","description":"A multi-view pose estimation dataset of tethered fruit flies (Drosophila melanogaster) behaving spontaneously on an air-supported ball, filmed by six synchronized cameras at 300 Hz (Karashchuk et al., Cell Reports 2021, the Anipose 3D tracking toolkit paper from the Tuthill lab). Each frame carries 30 keypoints, five joints (A to E, proximal to distal) on each of six legs. The labels are not hand-annotated: the Paninski Lab at Columbia University took filtered Anipose 3D predictions, dropped instances with mean reprojection error above 10 pixels, chose one instance per k-means cluster (25 per session) and reprojected the 3D poses to every camera view, masking keypoints with 2D error above 10 pixels. Contents: calibrations/ (per-session camera calibration .toml), labeled-data/ PNG frames, per-camera CollectedData CSVs for in-distribution and out-of-distribution sessions, videos, 51-frame clips for temporal smoothing, a Lightning Pose config and project.yaml; about 3.2 GB. Source data: Anipose repository on Dryad, DOI 10.5061/dryad.nzs7h44s4, CC0. Cite Karashchuk et al. 2021 and Lightning Pose.","displayTitle":"Fly Anipose: Six-Camera Drosophila Leg Keypoint Dataset at 300 Hz for Multi-View Pose Estimation (Lightning Pose, Columbia University)","huggingFaceUrl":"https://huggingface.co/datasets/paninski-lab/fly-anipose","id":"0f6f9e89-2ce0-5d97-9add-314dddc96823","license":"CC0-1.0","licenseEvidenceUrl":"https://datadryad.org/api/v2/datasets/doi%3A10.5061%2Fdryad.nzs7h44s4","objectCountEstimate":18675,"probeStatus":"ok","referenceUrl":"https://huggingface.co/datasets/paninski-lab/fly-anipose","region":"us","schemaCard":{"accessHints":[{"language":"python-datasets","snippet":"from datasets import load_dataset\nds = load_dataset(\"paninski-lab/fly-anipose\")"},{"language":"python-hub","snippet":"from huggingface_hub import snapshot_download\nsnapshot_download(repo_id=\"paninski-lab/fly-anipose\", repo_type=\"dataset\")"}],"caveats":["HuggingFace metadata: author=paninski-lab, license:cc-by-4.0, size_categories:n<1K, downloads=497"],"format":"mixed","generatedAt":"2026-11-21T20:48:47.894Z","layout":{"example":"CollectedData_Cam-A.csv","pattern":"{repoId}/{filename}"},"mountPath":"/mnt/archive","sampleHeaders":[{"bytes":"scorer,anipose,anipose,anipose,anipose,anipose,anipose,anipose,anipose,anipose,anipose,anipose,anipose,anipose,anipose,anipose,anipose,anipose,anipose,anipose,anipose,anipose,anipose,anipose,anipose,anipose,anipose,anipose,anipose,anipose,anipose,anipose,anipose,anipose,anipose,anipose,anipose,anipose,anipose,anipose,anipose,anipose,anipose,anipose,anipose,anipose,anipose,anipose,anipose,anipose,anipose,anipose,anipose,anipose,anipose,anipose,anipose,anipose,anipose,anipose,anipose\nbodyparts,L1A,L1A,L1B,L1B","encoding":"utf8","key":"CollectedData_Cam-A.csv","truncatedAt":512}],"sampleObjects":["CollectedData_Cam-A.csv","CollectedData_Cam-A_test.csv","CollectedData_Cam-B.csv","CollectedData_Cam-B_test.csv","CollectedData_Cam-C.csv"],"sidecars":[{"bytes":"---\ntask_categories:\n  - keypoint-detection\nlicense: cc-by-4.0\ntags:\n  - biology\n  - pose-estimation\n  - fly\n  - lightning-pose\npretty_name: Fly Anipose\nsize_categories:\n  - 1K<n<10K\n---\n\n# Fly Anipose — Lightning Pose Multi-View Dataset\n\nMulti-camera pose estimation dataset for fly leg keypoints, packaged for use with [Lightning Pose](https://github.com/paninski-lab/lightning-pose).\n\n## Dataset Description\n\nHead-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.\n\n> **Note:** Labels in this dataset are **not hand-labeled**. They are filtered Anipose 3D predictions reprojected to each camera view (see [Label Construction](#label-construction) below).\n\nSource data: Karashchuk et al., *Cell Reports* 2021 — original archive at https://doi.org/10.5061/dryad.nzs7h44s4.\n\n## Label Construction\n\nFor a subset of sessions in the Anipose data repository that contain 3D pose predictions, labels were constructed as follows:\n\n1. Remove any instance where average 3D reprojection error is > 10 pixels.\n2. Run k-means clustering on the remaining 3D poses (25 clusters per session) and select one instance per cluster.\n3. Use the filtered 2D reprojections (not the original predictions) as labels.\n4. Set any keypoint where per-keypoint 2D reprojection error > 10 pixels to NaN.\n\n## Data Splits\n\n| Split | Labeled instances | Sessions |\n|-------|---------------:|---------:|\n| In-distribution (InD) | 377 | 16 |\n| Out-of-distribution (OOD) | 300 | 12 |\n\nInD 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.\n\n- `CollectedData_{cam}.csv` — InD labels; `videos/` — InD videos\n- `CollectedData_{cam}_test.csv` — OOD labels; `videos_test/` — OOD videos\n\n## Keypoints\n\n30 keypoints per frame: 5 joints (A–E, proximal to distal) on each of 6 legs.\n\n| Left legs | Right legs |\n|-----------|------------|\n| L1A–L1E (foreleg) | R1A–R1E (foreleg) |\n| L2A–L2E (midleg)  | R2A–R2E (midleg)  |\n| L3A–L3E (hindleg) | R3A–R3E (hindleg) |\n\n## Directory Structure\n\n```\nfly-anipose/\n├── calibrations/                       # Per-session camera calibration files (.toml)\n├── labeled-data/                       # Extracted frames per session; includes ±2 context frames\n├── videos/                             # InD session video clips\n├── videos_test/                        # OOD session video clips\n├── videos-for-each-labeled-frame/      # 51-frame videos centered on each OOD labeled frame\n├── CollectedData_{cam}.csv             # InD 2D keypoint labels (x,y per keypoint) for each camera view\n├── CollectedData_{cam}_test.csv        # OOD 2D keypoint labels for each camera view\n├── config_fly-anipose.yaml             # Sample Lightning Pose training config\n└── project.yaml                        # View and keypoint definitions (required by LP App)\n```\n\nThe `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](https://github.com/paninski-lab/eks).\n\nSee the Lightning Pose documentation for full details on the [multi-view data directory structure](https://lightning-pose.readthedocs.io/en/latest/source/directory_structure_reference/multiview_structure.html).\n\n## Usage with Lightning Pose\n\nThe included `config_fly-anipose.yaml` is a ready-to-use training config. Key settings:\n\n- **Image resize:** 256 × 256\n- **Backbone:** `vits_dino`\n- **Keypoints:** 30\n- **Views:** Cam-A, Cam-B, Cam-C, Cam-D, Cam-E, Cam-F\n\nUpdate `data.data_dir` to absolute paths on your machine before training.\n\n```bash\nlitpose train config_fly-anipose.yaml\n```\n\n## Citation\n\nIf you use this dataset, please cite the o","encoding":"utf8","key":"README.md","name":"README.md","size":4564,"truncated":true}],"sourceObjectKey":"CollectedData_Cam-A.csv","starterCell":"import pandas as pd\ndf = pd.read_csv(\"hf://datasets/paninski-lab/fly-anipose/CollectedData_Cam-A.csv\")\ndf.head()","starterSamplePath":"CollectedData_Cam-A.csv","totalObjects":18675},"summary":"A multi-view pose estimation dataset of tethered fruit flies (Drosophila melanogaster) behaving spontaneously on an air-supported ball, filmed by six synchronized cameras at 300…","title":"Fly Anipose: Six-Camera Drosophila Leg Keypoint Dataset at 300 Hz for Multi-View Pose Estimation (Lightning Pose, Columbia University)","updatedAt":"2026-11-22T01:11:19.174Z","visibility":"public","websiteUrl":"https://github.com/paninski-lab/lightning-pose"}}