Marmoset3k — Lightning Pose Single-View Dataset

Top-down pose estimation dataset for common marmosets (Callithrix jacchus), packaged for use with Lightning Pose.

Dataset Description

Marmosets were recorded from a top-down camera in two configurations: alone (single) and in pairs. For pair sessions, individual animals are tracked — each labeled frame contains one focal animal cropped from a scene with two marmosets. 16 full-body keypoints are labeled per animal.

Source data: original archive at https://zenodo.org/records/14672425.

Data Splits

Split	Labeled frames	Sessions
In-distribution (InD)	3,920	4 (2 pair, 2 single)
Out-of-distribution (OOD)	824	2 (1 pair, 1 single)

InD and OOD sessions contain different animals.

• CollectedData.csv — InD labels; videos/ — InD videos
• CollectedData_test.csv — OOD labels

Keypoints

16 full-body keypoints.

Region	Keypoints
Head	head, leftear, rightear
Torso	neck, spinemid
Forelimbs	leftelbow, rightelbow, lefthand, righthand
Hindlimbs	leftknee, rightknee, leftfoot, rightfoot
Tail	tailbase, tailmid, tailend

Directory Structure

marmoset3k/
├── labeled-data/              # Extracted frames per session; includes ±2 context frames
├── videos/                    # InD session video clips
├── CollectedData.csv          # InD 2D keypoint labels (x,y per keypoint)
├── CollectedData_test.csv     # OOD 2D keypoint labels
├── config_marmoset3k.yaml     # Sample Lightning Pose training config
└── project.yaml               # View and keypoint definitions (required by LP App)

See the Lightning Pose documentation for full details on the single-view data directory structure.

Usage with Lightning Pose

The included config_marmoset3k.yaml is a ready-to-use training config. Key settings:

• Image resize: 256 × 256
• Backbone: resnet50_animal_ap10k
• Keypoints: 16

Update data.data_dir to an absolute path on your machine before training.

litpose train config_marmoset3k.yaml

Citation

If you use this dataset, please cite:

@article{cheng2025real,
  title     = {A real-time, multi-subject three-dimensional pose tracking system
               for the behavioral analysis of non-human primates},
  author    = {Cheng, Chaoqun and Huang, Zijian and Zhang, Ruiming and Huang, Guozheng
               and Wang, Han and Tang, Likai and Wang, Xiaoqin},
  journal   = {Cell Reports Methods},
  volume    = {5},
  number    = {2},
  year      = {2025},
  publisher = {Elsevier}
}

Original data archive: https://zenodo.org/records/14672425.