pretty_name: AniTrack Core-4
language:
- en
license: other
license_name: source-specific-terms
license_link: LICENSE.md
task_categories:
- keypoint-detection
- object-detection
tags:
- mouse
- rodent
- animal-pose
- core4
- multi-animal
- dataset-recipe
annotations_creators:
- expert-generated
- crowdsourced
AniTrack
The public data companion to
LuminBench-AniTrack:
mouse detection and Core-4 pose (nose, left_ear, right_ear, tail_base).
What is published here?
This initial release contains source-download manifests, semantic mappings,
an acquisition script, and the training-data recipe. It does not contain
image shards, per-image labels, pretrained weights, or a ready-to-load
datasets.load_dataset() training table. The counts below describe the original
prepared training views, not payloads currently hosted on this Hub repository.
Upstream images and labels must be acquired from their original publishers, under each source's terms. Public availability is not permission to relicense or mirror every source. See LICENSE.md.
Frozen recipe: Bridge-33K Rotating v2
| Partition / training view | Frames | Mouse annotations | Use |
|---|---|---|---|
| Pose training | 32,713 | 53,420 | Gradient updates |
| Detector training | 32,754 | 53,502 | Gradient updates |
| External validation | 2,662 | 4,801 | Checkpoint selection |
| DLC historical test | 263 | 285 boxes / 283 Core-4-evaluable mice | Final comparison only |
The union before cleaning contains 27,826 external training frames plus 5,114 TopViewMouse training frames. The latter includes 1,012 Openfield training images; the disjoint 54 Openfield test images remain inside the frozen 263-image historical test. Historical test images must not be used to train, generate training pseudo-labels, select checkpoints, or tune thresholds.
Source families
Counts are upstream human-supervised frames across their available splits, before the AniTrack split and cleaning operations.
| Source | Frames | Supervision / view | Terms and evidence |
|---|---|---|---|
| TopViewMouse5K v2 | 5,114 train + 263 test | Human 27-point union; partial Core-4; mostly top view | Constituent-source terms; no single annotator-count claim |
| MARS | 15,000 | Two mice, seven points, top view | CC-BY-NC-4.0; five-worker median labels |
| Kumar OFA | 8,910 | One mouse, twelve points, top view | Custom non-commercial terms; human labels, rater count unreported |
| Lightning Pose CRIM13 | 5,260 labeled centers | Two mice, seven points, top view | CC-BY-4.0; five-annotator median; 21,040 neighboring frames are not human-labeled centers |
| SLEAP mice_hc | 1,474 | Two mice, five points, overhead | Human labels; preserve dataset credits and verify redistribution terms |
| SLEAP mice_of | 1,000 | One to five mice, partial Core-4, below-floor | Human labels; not a strict-overhead source; verify redistribution terms |
| Mouse Lockbox | 544 usable top-camera rows | One mouse, apparatus occlusions | CC-BY-4.0; human pose labels; only verified top camera used |
Sampling, cleaning, and limitations
MARS, Kumar and CRIM13 are capped at 1,242 frames per epoch each, using seeded rotating windows. Smaller sources retain their eligible frames. Pose training rotates the selected target in multi-mouse images; detector training retains all valid boxes in a selected frame. The original pose virtual epoch contains 11,070 crops, with 600 epochs and global batch 64 (103,800 updates).
Human partial keypoints are masked, never filled with teacher predictions. Near-complete overlap is excluded from the isolated pose-crop view but retained for detector supervision when boxes remain useful. Empty records and exact duplicate annotations are removed. The original detector loader additionally skipped seven invalid normalized-label records; do not silently repair the published recipe and still call it an exact reproduction.
Published video/group splits are preserved where available. MARS lacks released session identifiers; its deterministic frame-hash split is not proof of session-independent generalization. TopViewMouse does not expose reliable session IDs either. SLEAP mice_of is a below-floor domain and should be reported separately in source-level evaluations.
Acquire the original archives
Clone this small metadata repository on the machine where you intend to store the data:
git clone https://huggingface.co/datasets/LuminScience/AniTrack
cd AniTrack
python3 download_sources.py --list
# Only after reviewing the upstream terms; downloads are explicit:
python3 download_sources.py --download --accept-source-terms \
--output-root /your/data/raw
The script uses only Python's standard library, validates pinned publisher MD5
checksums (and sizes where available), and does not extract or alter the source
archives. Existing valid files are reused; invalid existing files cause a hard
failure. Use --sources mars_topview_pose to acquire a selected source. No
training split is created by this acquisition step.
The manifests/ files preserve native layouts and Core-4 mappings. The external
preparation configuration records the split and missing-label policies;
recipe.json records the combined training contract. Conversion, split audits,
and pose/detector training live in the companion code repository. Pin revisions
of both repositories for any experiment. The historical external staging
manifest filename containing gated refers to two now-directly-downloadable
sources, not a Hub access gate.
Scope
This is a frame-level detection/pose resource, not newly annotated ground truth for persistent tracking identities. No new animal experiments were conducted to produce this collection. Cite the original datasets and respect their animal research, attribution, and usage requirements. Candidate datasets awaiting access or semantic review are not counted in this recipe.