Dataset Viewer
Auto-converted to Parquet Duplicate
Search is not available for this dataset
video
video
End of preview. Expand in Data Studio

SLAP-2M

Multi-camera 3D pose of one to four mice in an arena with graded levels of environmental enrichment, released with LUC3D (https://luc3d.sleap.ai), a browser-based tool for 3D multi-animal labeling, cross-view tracking, triangulation and proofreading.

Sessions 74, 1 to 60 minutes each
Cameras 8, synchronized, 1280Γ—1024, grayscale
Frame rate 30 fps, except the 13 sessions of 2022-10-19, which are 60 fps
Frames 1,954,440 per camera, 15.6 million video frames in total
Poses 2.81 million 3D poses, proofread in 3D
Animals 1 to 4 per session (32, 35, 4 and 3 sessions with 1, 2, 3 and 4 mice): eight CD-1, four C57BL/6J and four BALB/c mice
Difficulty rating 1 to 7, from the number of animals and the enrichment objects
Skeleton 15 nodes, the same as MouseDyad-10M (mouse_skeleton.toml)

The data come from the Falkner lab (Princeton).

Layout

data_sheet.csv                 one row per session: animals and coat colors, enrichment, bedding, frames, difficulty and the path of every file
mouse_skeleton.toml, cage_skeleton.toml, cube_skeleton.json
<date>/<session>/
β”œβ”€β”€ points3d.h5                3D poses: tracks (frames, animals, 15, 3), mm, calibration frame
β”œβ”€β”€ aligned_points3d.h5        the same, aligned to the arena (alignment.toml)
β”œβ”€β”€ aligned_cage_points3d.h5   the arena landmarks in the aligned frame
β”œβ”€β”€ calibration.toml           camera intrinsics, distortion and extrinsics
β”œβ”€β”€ alignment.toml
β”œβ”€β”€ reprojections.h5           the 3D reprojected into each camera: <cam> = (frames, animals, 15, 2)
└── <cam>/                     back backL mid midL side sideL top topL
    β”œβ”€β”€ <cam>-<ts>-0000_h264_CRF20_denoised.mp4                      the video
    β”œβ”€β”€ *.predictions.slp                                            SLEAP predictions before proofreading
    β”œβ”€β”€ *.predictions.proofread.slp (+ .analysis.h5)                 proofread 2D (not for side and sideL)
    └── *.reprojections.slp (+ .h5)                                  the 3D reprojected, as SLEAP labels

Some sessions also hold working files from the original tracking: *_CRF12_denoised.mp4.predictions.static.externals.slp (36 sessions) and *.npy (12 sessions).

The .slp and .h5 names mention CRF30 or CRF12 from the original encodes. Every .slp opens the _h264_CRF20_denoised.mp4 in its own folder, which is the only video distributed.

import h5py, sleap_io as sio
with h5py.File("2022-10-07/10072022120554/aligned_points3d.h5") as f:
    pts3d = f["tracks"][:]  # (frames, animals, 15, 3), mm
labels = sio.load_slp("2022-10-07/10072022120554/back/back-10072022120551-0000_h265_CRF30_denoised.mp4.predictions.proofread.slp")

Videos

The original recordings are h265 at CRF 12, 709 GB in total. The videos here are re-encoded to x264, CRF 20, preset medium, GOP 250, yuv420p 8-bit (189 GB):

  • every video has the same frame count as its source;
  • pixel values agree with the source to a mean absolute difference of about 1 gray level;
  • each source's color range is kept. The 10-bit 2022-10-30 sources become 8-bit.

In tests, running the pose models on re-encoded frames instead of lossless frames moved keypoints by a median 0.3 to 2.1 pixels, about the noise floor of a near-lossless re-encode.

Known issues

  • side and sideL have no proofread 2D.
  • Six sessions (10072022142111 to 10072022151549) have a blank duration in data_sheet.csv. Their frame counts are correct.
  • Heights up to 393 mm and floor dips down to βˆ’42 mm occur only in difficulty 5–7 sessions with enrichment objects, consistent with climbing.

Citation

Leonardis, E. J., Oline, S. N., … Falkner, A. L.* & Pereira, T. D.* LUC3D: Multi-animal 3D pose tracking in the browser (in submission, 2026).

Code for every figure and analysis in the paper: https://github.com/talmolab/figures-luc3d.

Downloads last month
885