juancobos/MultiTaskMouseBehaviour

barnes_maze	direct_interaction	marble
nort	nort2	open_field
social_interaction	t_maze	three_chamber

Six-second clips, annotations overlaid (boxes, masks, keypoints, track ids).

Multi-Task Mouse Behaviour Dataset

Bounding boxes, instance masks, 27-point poses and persistent track ids for laboratory mice, on the same frames, across nine standard behavioural assays. 14,300 frames, 27,716 annotated instances, 1280×720.

Most animal-behaviour datasets give you one annotation type. Training or evaluating a model that does detection and segmentation and pose and tracking usually means stitching together sources that disagree on species, viewpoint, and label conventions. This dataset provides all four on every frame, under one category schema, from real recordings of running experiments.

Where the annotations come from

Every label is machine-generated by the DeepLabSAM pipeline. The pipeline is a two-stage composition:

1. Detection, segmentation and tracking — SAM 3 (video). A text prompt ("mouse") drives promptable detection and tracking, producing a pixel-accurate mask and a persistent track_id per animal per frame.
2. Pose — DeepLabCut SuperAnimal (superanimal_topviewmouse), mask-gated. Each instance is cropped to its own box and the crop's background is zeroed using that instance's mask before it reaches the pose head. Masking is applied after normalisation, where the network's neutral value is 0, so the background drops to that baseline instead of becoming an out-of-distribution black box.

Mask-gating is what makes stage 2 usable as an annotator. A top-down pose model fed a raw crop of two touching animals has no way to know which one it should fit; gated by the mask, the intended animal is the only signal present. That is why the poses survive the crowded assays (open_field, social_interaction) where a plain detect-then-crop cascade puts keypoints on the neighbour.

Tasks

Each task is one continuous recording of a standard behavioural assay, filmed from above and decoded to consecutive frames at 30 fps. Frames are contiguous, so neighbours are near-duplicates — the ~14k frames are roughly 7.9 minutes of footage, not 14k independent samples.

Task	Frames	Instances	Animals	≈ Duration	Median box (px²)
barnes_maze	1,847	1,847	1	62 s	8,241
direct_interaction	1,430	2,860	2	48 s	92,987
marble	1,797	1,797	1	60 s	102,820
nort	1,719	1,708	1	57 s	2,880
nort2	1,053	1,053	1	35 s	22,518
open_field	1,939	7,690	4	65 s	1,377
social_interaction	1,472	5,888	4	49 s	2,352
t_maze	2,128	2,128	1	71 s	1,409
three_chamber	915	2,745	3	30 s	2,816
Total	14,300	27,716		≈ 7.9 min

barnes_maze — Spatial-learning assay. One mouse on a circular platform ringed with escape holes.

direct_interaction — Two mice in a small bedding-filled arena with no divider between them. The only task with two coat colours.

marble — Marble-burying assay for repetitive and anxiety-like behaviour.

nort / nort2 — Novel Object Recognition: one mouse investigating objects in an open arena, recorded twice in different arenas. The only within-assay pair.

open_field — Locomotion and anxiety assay. Four mice in a quadrant-divided arena, one per compartment.

social_interaction — Four mice in a divided arena, approaching and contacting one another across dividers.

t_maze — Spatial working-memory assay. One mouse running the arms of a T-shaped maze.

three_chamber — Sociability assay. Three mice in a chamber divided by transparent partitions.

Four of nine tasks are multi-animal.

Layout

<task>/
  images/frame_00000.jpg ...        # consecutive decoded frames, 1280x720
  annotations.json                  # COCO: bbox, segmentation, track_id, keypoints
assets/gifs/<task>.gif              # gallery previews for this card
assets/examples/<task>.mp4          # 10 s unannotated clips, e.g. as demo inputs

One COCO file per task. Splits and training exports (e.g. RF-DETR layout) are built from these with the mtmb package.

Annotation schema

Standard COCO detection + keypoints, one category (mouse, id 1), with a few additions:

Field	Meaning
track_id	Persistent identity across frames within a task.
score	SAM 3's detection confidence.
keypoints	Flat [x, y, v] × 27, SuperAnimal topviewmouse order.
keypoint_scores	Per-keypoint model confidence — use this to re-threshold.
num_keypoints	How many cleared the 0.3 threshold.
images[*].frame_index	Position in the source recording (temporal order).

The 27 keypoints follow SuperAnimal topviewmouse naming exactly, so DeepLabCut models are directly comparable:

nose, left_ear, right_ear, left_ear_tip, right_ear_tip, left_eye, right_eye, neck, mid_back, mouse_center, mid_backend, mid_backend2, mid_backend3, tail_base, tail1, tail2, tail3, tail4, tail5, left_shoulder, left_midside, left_hip, right_shoulder, right_midside, right_hip, tail_end, head_midpoint

License

CC BY-NC 4.0. The pose annotations are the output of DeepLabCut's superanimal_topviewmouse checkpoint, whose weights (distinct from the LGPL-3.0 DeepLabCut codebase) are licensed for academic, non-commercial use only. That restriction is what this dataset inherits and passes on; it is not a choice made independently of the pipeline that produced the labels.

Acknowledgments

Annotations were produced with:

• SAM 3 (Meta AI / FAIR) — detection, segmentation and tracking.
• DeepLabCut SuperAnimal, superanimal_topviewmouse (Mathis Lab, EPFL) — pose estimation. See Ye et al., 2024.

Recordings were collected at IGF, CNRS and UPV/EHU.

Citation

A paper describing this dataset is in preparation. Citation details (BibTeX) will be added here on publication.

@dataset{mtmb,
  title   = {Multi-Task Mouse Behaviour Dataset},
  author  = {TBD},
  year    = {TBD},
  note    = {TBD},
}