CalMS21-mini

A small subset of the Caltech Mouse Social Interactions (CalMS21) dataset, packaged for teaching. It contains the first 40 sequences of the CalMS21 Task 1 training set (pose tracking plus frame-level behavior annotations) in a single compressed NumPy file, so a classroom of students can download it quickly in Colab without fetching the full ~460 MB archive.

It was made for the tutorial Extracting Features and Detecting Actions in Pose Datasets (Kennedy Lab, Scripps Research). For research use, please download the full dataset from CaltechDATA.

Dataset description

The sequences are top-view recordings of pairs of freely behaving mice in a standard resident-intruder assay. Both animals were tracked with MARS, and each frame was manually annotated by a single expert annotator for one of four behaviors.

Sequences	40 (from CalMS21 Task 1, train split)
Frames	242846
Frame rate	30 fps
Animals per frame	2 (index 0 = resident, index 1 = intruder)
Keypoints per animal	7: nose, left ear, right ear, neck, left hip, right hip, tail base
Behavior labels	attack (0), investigation (1), mount (2), other (3)
Annotators	1
File	calms21_task1_train_subset40.npz (23.2 MB)

File format

calms21_task1_train_subset40.npz holds, for each sequence i = 0 … 39:

Key	Shape	dtype	Contents
keypoints_{i}	(frames, 2, 2, 7)	float32	pixel coordinates: (frame, mouse, x/y, keypoint)
scores_{i}	(frames, 2, 7)	float16	keypoint confidence from MARS, 0 (worst) to 1 (best)
annotations_{i}	(frames,)	int8	behavior label per frame (see vocabulary above)

plus two shared entries:

Key	Contents
names	original CalMS21 sequence names, e.g. task1/train/mouse001_task1_annotator1, in original file order
metadata_json	JSON list with each sequence's original metadata dict (annotator id and behavior vocabulary)

Usage

import json, urllib.request
import numpy as np

url = 'https://huggingface.co/datasets/annkennedy/calms21-mini/resolve/main/calms21_task1_train_subset40.npz'
urllib.request.urlretrieve(url, 'calms21_mini.npz')

with np.load('calms21_mini.npz') as d:
    names = [str(n) for n in d['names']]
    metadata = json.loads(str(d['metadata_json']))
    keypoints = d['keypoints_0']        # (frames, 2, 2, 7)
    annotations = d['annotations_0']    # (frames,)
    vocab = metadata[0]['vocab']        # {'attack': 0, 'investigation': 1, 'mount': 2, 'other': 3}

The tutorial rebuilds these arrays into the same nested-dictionary layout as the original CalMS21 JSON (data['annotator-id_0'][name] → keypoints, scores, annotations, metadata).

Splits

These are not the official CalMS21 train/test splits: all 40 sequences come from the official training set. Report results on the official CalMS21 test set when comparing against published benchmarks.

Intended use and limitations

• Intended for teaching and quick prototyping of pose-based behavior analysis and classification.
• Covers only 40 sequences from a single annotator and a single assay, so models trained on it will not generalize well and results should not be compared with benchmarks on the full dataset.
• Behaviors are imbalanced: other and investigation make up most frames, and attack appears in only a minority of sequences.
• Raw video, unlabeled sequences, and the Task 2 and Task 3 data are not included.

Source and license

Derived from the CalMS21 dataset (doi:10.22002/D1.1991), released under CC BY 4.0. This subset is redistributed under the same license. Changes from the original: selection of the first 40 Task 1 training sequences, conversion from JSON to compressed NumPy, and reduced numeric precision as described above.

Citation

If you use this data, please cite the original CalMS21 paper:

@article{sun2021multi,
  title={The multi-agent behavior dataset: Mouse dyadic social interactions},
  author={Sun, Jennifer J and Karigo, Tomomi and Chakraborty, Dipam and Mohanty, Sharada P and Wild, Benjamin and Sun, Quan and Chen, Chen and Anderson, David J and Perona, Pietro and Yue, Yisong and Kennedy, Ann},
  journal={Advances in neural information processing systems},
  volume={2021},
  number={DB1},
  pages={1},
  year={2021}
}