healthonrails/sleap_mice_hc_yolo_pose

mice_hc — Home-cage mice pose (SLEAP) converted to YOLO pose format with Annolid (https://github.com/healthonrails/annolid)

Dataset Summary

mice_hc is a two-animal pose dataset consisting of pairs of male and female white Swiss Webster mice recorded from an overhead home-cage view with light bedding. The animals are low contrast relative to background, which makes it a useful benchmark for robust pose estimation in challenging conditions.

This Hugging Face dataset provides the original split labels (Train/Val/Test) from SLEAP .pkg.slp files converted to YOLO pose format using Annolid.

Key facts:

• —Videos: 40
• —Frame rate: 40 FPS
• —Image size: 1280 × 1024 × 1 (grayscale)
• —Resolution: 1.9 px/mm
• —Skeleton: 5 keypoints (“5 nodes”)
• —# Animals per frame: 2
• —Identity tracking: ❌ (no consistent identity labels across frames)
• —Labels: 1474 frames, 2948 instances

Original artifacts (SLEAP format):

• —Train: train.pkg.slp
• —Validation: val.pkg.slp
• —Test: test.pkg.slp (See “Source data” below.)

Supported Tasks

• —Keypoint detection / pose estimation (multi-animal) in a single view (top-down).
• —Training YOLO-style pose models (e.g., Ultralytics YOLO pose).

Data Format (YOLO Pose)

This dataset is exported in YOLO pose label format (compatible with common YOLO pose toolchains).

Directory layout

sleap_mice_hc_yolo_pose/
  images/
    train/
    val/
    test/
  labels/
    train/
    val/
    test/
  data.yaml
  pose_schema.json

Label file format

Each image has a corresponding text file in labels/<split>/. Each line corresponds to one mouse instance:

<class_id> <x_center> <y_center> <width> <height> <x1> <y1> <v1> ... <xK> <yK> <vK>

Where:

• —coordinates are normalized to [0,1] by image width/height
• —K = 5 keypoints
• —v is keypoint visibility (commonly: 0=not labeled, 1=labeled but not visible/occluded, 2=visible; exact convention depends on your training code—see “Notes on visibility”)

data.yaml example

path: mice_hc_yolo
train: images/train
val: images/val
test: images/test

nc: 1
names: ["mouse"]

kpt_shape: [5, 3]
# optional (some trainers support explicit edges):
# skeleton:
#   - [0, 1]
#   - [0, 2]
#   ...

Dataset Splits

The dataset preserves the original SLEAP random split:

• —train
• —val
• —test

Note: Identity is not provided, so although there are typically 2 mice per frame, instances are treated independently.

Pose Schema

The original dataset uses a 5-node skeleton. The exact keypoint names depend on the provided schema used during conversion. This repo includes a pose_schema.json used by Annolid to define:

• —keypoint order
• —optional symmetry pairs
• —optional edges

Example skeleton schema structure:

How the Conversion Was Done (SLEAP → YOLO Pose) with Annolid

Source data

This dataset is derived from the SLEAP sample/benchmark dataset “mice_hc”:

• —Train: https://storage.googleapis.com/sleap-data/datasets/eleni_mice/random_split1/train.pkg.slp
• —Validation: https://storage.googleapis.com/sleap-data/datasets/eleni_mice/random_split1/val.pkg.slp
• —Test: https://storage.googleapis.com/sleap-data/datasets/eleni_mice/random_split1/test.pkg.slp
• —Example clip: https://storage.googleapis.com/sleap-data/datasets/eleni_mice/clips/20200111_USVpairs_court1_M1_F1_top-01112020145828-0000%400-2560.mp4
• —Example tracking: https://storage.googleapis.com/sleap-data/datasets/eleni_mice/clips/20200111_USVpairs_court1_M1_F1_top-01112020145828-0000%400-2560.slp

Conversion steps (recommended workflow)

1. 1.Download SLEAP `.pkg.slp` files for train/val/test.
2. 2.Extract frames + annotations from .pkg.slp.
3. 3.Compute per-instance bounding boxes from keypoints (or SLEAP instance extents).
4. 4.Write YOLO pose labels (normalized bbox + normalized keypoints).
5. 5.Save `data.yaml` and the pose_schema.json.

Notes on visibility (v)

SLEAP annotations may not always map 1:1 to YOLO’s visibility conventions. In this conversion:

• —labeled keypoints are typically exported with v=2
• —missing/unlabeled points may be v=0

If you need strict COCO-style handling (e.g., occluded vs visible), you may need to add a rule based on SLEAP point confidence/visibility metadata (if present).

Intended Uses

• —Training and benchmarking multi-animal pose estimation models in challenging low-contrast home-cage settings.
• —Testing robustness of detectors + keypoint heads under occlusion and interaction.

Limitations

• —No consistent identity labels across frames (two animals per frame but not tracked by ID).
• —Overhead grayscale imagery with low contrast may require careful augmentation and/or background normalization.

Citation

Please cite the original SLEAP paper and dataset contributors:

• —Pereira et al. (2022), Nature Methods
• —Eleni Papadoyannis, Mala Murthy, Annegret Falkner

@article{yang2023automated, title={Automated Behavioral Analysis Using Instance Segmentation}, author={Yang, Chen and Forest, Jeremy and Einhorn, Matthew and Cleland, Thomas A}, journal={arXiv preprint arXiv:2312.07723}, year={2023} }

@misc{yang2020annolid, author = {Chen Yang and Jeremy Forest and Matthew Einhorn and Thomas Cleland}, title = {Annolid: an instance segmentation-based multiple animal tracking and behavior analysis package}, howpublished = {\url{https://github.com/healthonrails/annolid}}, year = {2020} }

License / Terms

This Hugging Face dataset repo contains converted annotations and (optionally) extracted frames derived from the original SLEAP dataset files hosted at the URLs above. Please ensure your redistribution and usage comply with the original dataset’s license/terms and any institutional requirements.