EpicPinkPenguin/cmu_panoptic_dataset_instance_segmentation_mask

CMU Panoptic person instance masks (SAM 3.1 pseudo-labels)

TODO before making this repo public: confirm permission from the CMU Panoptic Studio organisers. Until then the repo must stay gated and private. See LICENSE_NOTICE.md.

Per-person instance segmentation masks for the CMU Panoptic Studio dataset, generated with SAM 3.1 (facebook/sam3.1 @ daa63191845a41281374e725f4c9e51c7a824460) and aligned to the dataset's own 3D skeletons: masks and pose share the same person id across all five VoxelPose HD cameras (3, 6, 12, 13, 23) and all frames.

These are model-generated pseudo-labels, not human annotations. Every published frame carries automated QA metrics (index.parquet); read the numbers below before trusting the masks for your use case.

This repo contains no CMU pixels -- masks, metadata and code only. Videos must be obtained directly from CMU under CMU's terms: run scripts/download_videos.py.

Layout

MANIFEST.json                        # full provenance: checkpoints, config, offsets, checksums
availability.json                    # probe results for all Tier A sequences
index.parquet                        # one row per (sequence, camera, frame): QA + byte offsets
masks/<seq>/hd_<cam>/instances.jsonl.zst   # per-frame COCO RLE instances
masks/<seq>/calib_subset.json        # the 5 HD cameras, verbatim from CMU calibration
scripts/download_videos.py           # fetch the source videos from CMU into a local cache
scripts/loader.py                    # yields (image, pose3d, instances) joined on person id
LICENSE_NOTICE.md

Each line of instances.jsonl.zst is one frame: person_id is the Panoptic id (identical in every camera); negative ids are text-detector instances with no 3D counterpart (no_3d_gt flag) -- real people outside the capture volume, not noise. rle=null means the person is reconstructed in 3D but not visible in this view. depth_proxy is camera-to-person distance in cm, giving free occlusion ordering.

Method

SAM 3.1's video predictor runs with a person text prompt in chunks of 600 frames with 10-frame overlaps; tracks are stitched across chunk seams by mask IoU (SAM ids are session-local). Panoptic person ids are then assigned by fitting whole trajectories of projected 3D joints against the tracked masks, with depth-weighted disambiguation for contested masks. The video/pose frame offset is resolved empirically per sequence (see MANIFEST.json), not assumed.

Coverage

Tier A of the corpus: 65 probe-verified sequences with HD pose, of which the following are currently packaged:

sequence	category	cameras	frames per camera
160224_haggling1	haggling	5	hd_12: 8716, hd_13: 8716, hd_23: 8716, hd_03: 8716, hd_06: 8716
160226_haggling1	haggling	5	hd_12: 11465, hd_13: 11465, hd_23: 11465, hd_03: 11465, hd_06: 11465
160422_haggling1	haggling	5	hd_12: 13580, hd_13: 13580, hd_23: 13580, hd_03: 13580, hd_06: 13580
160422_ultimatum1	ultimatum	5	hd_12: 26794, hd_13: 26794, hd_23: 26794, hd_03: 26794, hd_06: 26794
160906_band2	musical_instruments	5	hd_12: 7362, hd_13: 7362, hd_23: 7362, hd_03: 7362, hd_06: 7362
160906_ian1	toddler	5	hd_12: 2847, hd_13: 2847, hd_23: 2847, hd_03: 2847, hd_06: 2847
160906_ian3	toddler	5	hd_12: 7368, hd_13: 7368, hd_23: 7368, hd_03: 7368, hd_06: 7368
160906_ian5	toddler	5	hd_12: 2872, hd_13: 2872, hd_23: 2872, hd_03: 2872, hd_06: 2872
160906_pizza1	special_events	5	hd_12: 6582, hd_13: 6582, hd_23: 6582, hd_03: 6582, hd_06: 6582
161029_build1	special_events	5	hd_12: 8860, hd_13: 8860, hd_23: 8860, hd_03: 8860, hd_06: 8860
161029_car1	special_events	5	hd_12: 3525, hd_13: 3525, hd_23: 3525, hd_03: 3525, hd_06: 3525
161029_piano1	musical_instruments	5	hd_12: 8920, hd_13: 8920, hd_23: 8920, hd_03: 8920, hd_06: 8920
161029_tools1	special_events	5	hd_12: 5470, hd_13: 5470, hd_23: 5470, hd_03: 5470, hd_06: 5470
161202_haggling1	haggling	5	hd_12: 10850, hd_13: 10850, hd_23: 10850, hd_03: 10850, hd_06: 10850
170221_haggling_b1	haggling	5	hd_12: 14086, hd_13: 14086, hd_23: 14086, hd_03: 14086, hd_06: 14086
170221_haggling_b2	haggling	5	hd_12: 14266, hd_13: 14266, hd_23: 14266, hd_03: 14266, hd_06: 14266
170221_haggling_b3	haggling	5	hd_12: 14914, hd_13: 14914, hd_23: 14914, hd_03: 14914, hd_06: 14914
170221_haggling_m1	haggling	5	hd_12: 14626, hd_13: 14626, hd_23: 14626, hd_03: 14626, hd_06: 14626
170221_haggling_m2	haggling	5	hd_12: 14597, hd_13: 14597, hd_23: 14597, hd_03: 14597, hd_06: 14597
170221_haggling_m3	haggling	5	hd_12: 7074, hd_13: 7074, hd_23: 7074, hd_03: 7074, hd_06: 7074
170224_haggling_a2	haggling	5	hd_12: 19181, hd_13: 19181, hd_23: 19181, hd_03: 19181, hd_06: 19181
170224_haggling_a3	haggling	5	hd_12: 18132, hd_13: 18132, hd_23: 18132, hd_03: 18132, hd_06: 18132
170224_haggling_b1	haggling	5	hd_12: 19481, hd_13: 19481, hd_23: 19481, hd_03: 19481, hd_06: 19481
170224_haggling_b2	haggling	5	hd_12: 19181, hd_13: 19181, hd_23: 19181, hd_03: 19181, hd_06: 19181
170224_haggling_b3	haggling	5	hd_12: 18366, hd_13: 18366, hd_23: 18366, hd_03: 18366, hd_06: 18366
170228_haggling_a1	haggling	5	hd_12: 19181, hd_13: 19181, hd_23: 19181, hd_03: 19181, hd_06: 19181
170228_haggling_a2	haggling	5	hd_12: 18881, hd_13: 18881, hd_23: 18881, hd_03: 18881, hd_06: 18881
170228_haggling_a3	haggling	5	hd_12: 17443, hd_13: 17443, hd_23: 17443, hd_03: 17443, hd_06: 17443
170228_haggling_b1	haggling	5	hd_12: 23707, hd_13: 23707, hd_23: 23707, hd_03: 23707, hd_06: 23707
170228_haggling_b2	haggling	5	hd_12: 24037, hd_13: 24037, hd_23: 24037, hd_03: 24037, hd_06: 24037
170228_haggling_b3	haggling	5	hd_12: 21160, hd_13: 21160, hd_23: 21160, hd_03: 21160, hd_06: 21160
170307_dance5	dance	5	hd_12: 11617, hd_13: 11617, hd_23: 11617, hd_03: 11617, hd_06: 11617
170404_haggling_a1	haggling	5	hd_12: 9471, hd_13: 9471, hd_23: 9471, hd_03: 9471, hd_06: 9471
170404_haggling_a2	haggling	5	hd_12: 9675, hd_13: 9675, hd_23: 9675, hd_03: 9675, hd_06: 9675
170404_haggling_a3	haggling	5	hd_12: 11863, hd_13: 11863, hd_23: 11863, hd_03: 11863, hd_06: 11863
170404_haggling_b1	haggling	5	hd_12: 19062, hd_13: 19062, hd_23: 19062, hd_03: 19062, hd_06: 19062
170404_haggling_b2	haggling	5	hd_12: 19277, hd_13: 19277, hd_23: 19277, hd_03: 19277, hd_06: 19277
170404_haggling_b3	haggling	5	hd_12: 18558, hd_13: 18558, hd_23: 18558, hd_03: 18558, hd_06: 18558
170407_haggling_a1	haggling	5	hd_12: 14087, hd_13: 14087, hd_23: 14087, hd_03: 14087, hd_06: 14087
170407_haggling_a2	haggling	5	hd_12: 13847, hd_13: 13847, hd_23: 13847, hd_03: 13847, hd_06: 13847
170407_haggling_a3	haggling	5	hd_12: 12768, hd_13: 12768, hd_23: 12768, hd_03: 12768, hd_06: 12768
170407_haggling_b1	haggling	5	hd_12: 19122, hd_13: 19122, hd_23: 19122, hd_03: 19122, hd_06: 19122
170407_haggling_b2	haggling	5	hd_12: 18882, hd_13: 18882, hd_23: 18882, hd_03: 18882, hd_06: 18882
170407_haggling_b3	haggling	5	hd_12: 16364, hd_13: 16364, hd_23: 16364, hd_03: 16364, hd_06: 16364
170407_office2	special_events	5	hd_12: 5216, hd_13: 5216, hd_23: 5216, hd_03: 5216, hd_06: 5216
170915_office1	special_events	5	hd_12: 5378, hd_13: 5378, hd_23: 5378, hd_03: 5378, hd_06: 5378
170915_toddler5	toddler	5	hd_12: 3574, hd_13: 3574, hd_23: 3574, hd_03: 3574, hd_06: 3574
171026_cello3	musical_instruments	5	hd_12: 1728, hd_13: 1728, hd_23: 1728, hd_03: 1728, hd_06: 1728
171026_pose1	range_of_motion	5	hd_12: 22466, hd_13: 22466, hd_23: 22466, hd_03: 22466, hd_06: 22466
171026_pose2	range_of_motion	5	hd_12: 14974, hd_13: 14974, hd_23: 14974, hd_03: 14974, hd_06: 14974
171026_pose3	range_of_motion	5	hd_12: 7181, hd_13: 7181, hd_23: 7181, hd_03: 7181, hd_06: 7181
171204_pose1	range_of_motion	5	hd_12: 27561, hd_13: 27561, hd_23: 27561, hd_03: 27561, hd_06: 27561
171204_pose2	range_of_motion	5	hd_12: 37751, hd_13: 37751, hd_23: 37751, hd_03: 37751, hd_06: 37751
171204_pose3	range_of_motion	5	hd_12: 8920, hd_13: 8920, hd_23: 8920, hd_03: 8920, hd_06: 8920
171204_pose4	range_of_motion	5	hd_12: 31397, hd_13: 31397, hd_23: 31397, hd_03: 31397, hd_06: 31397
171204_pose5	range_of_motion	5	hd_12: 26902, hd_13: 26902, hd_23: 26902, hd_03: 26902, hd_06: 26902
171204_pose6	range_of_motion	5	hd_12: 22886, hd_13: 22886, hd_23: 22886, hd_03: 22886, hd_06: 22886

Excluded sequences and why:

• 160906_band3: known data corruption; dropped by TEMPO and prior work
• 170224_haggling_a1: pose coverage 0.00% below 90% threshold (0 pose frames vs ~21600 video frames for 12.0 min)

QA (automated, per activity category)

Joint containment = fraction of projected 3D joints landing inside their own person's mask. Temporal IoU = frame-to-frame mask IoU per track. Hull IoU = reprojection IoU of the visual hull carved from the five views. Full per-frame values are in index.parquet.

category	cameras	frames	masks	containment	temporal IoU	hull IoU	id switches	count jumps
dance	5	58085	81255	0.998	0.907	0.848	0	0
haggling	165	2629500	6663672	0.953	0.945	0.729	412	1425
musical_instruments	15	90050	188758	0.889	0.974	0.723	3	7
range_of_motion	45	1000190	1209746	0.998	0.961	0.774	15	23
special_events	30	175155	375902	0.902	0.951	0.707	3	9
toddler	20	83305	175198	0.916	0.931	0.719	3	19
ultimatum	5	133970	612248	0.914	0.946	0.598	29	41

Known limitations

• Pseudo-labels: no human ever verified an individual mask. Raw containment dips on steep views are dominated by occlusion (occlusion-adjusted containment is reported in per-sequence QA), but genuine mask errors exist.
• Negative-id bystander tracks fragment across chunk seams; their ids are not stable over time. Subject ids (>= 0) are.
• Source videos contain defects (e.g. corrupt GOPs decoding to solid green); frames flagged count_jump in index.parquet deserve suspicion.
• Coverage is limited to the Tier A sequences listed above.

Citation

If you use this data, cite the CMU Panoptic Studio papers (required by CMU) and SAM 3: Joo et al., Panoptic Studio: A Massively Multiview System for Social Motion Capture, ICCV 2015; Joo et al., TPAMI 2017; Simon et al., Hand Keypoint Detection in Single Images using Multiview Bootstrapping, CVPR 2017. Full BibTeX in LICENSE_NOTICE.md.