The Dataset Viewer has been disabled on this dataset.

YCBMultiTrack

A dynamic multi-object RGB-D benchmark for 6D pose tracking under heavy occlusion, with motion-capture ground truth. Released with Point2Pose (ECCV 2026). The synthetic data generator used to render the Sim split is open source as well.

Objects are picked up, moved, rotated, mutually occluded, and repeatedly carried fully out of frame by a human hand. Every frame carries per-object 6D pose, an instance mask, and two visibility labels.

Split Sequences Frames Objects Camera Archive
Real 11 16,314 5 YCB Intel RealSense D435i + OptiTrack 8.95 GB
Sim 5 2,705 7 YCB Isaac Sim, 4 environments 1.11 GB
Models β€” β€” 10 YCB meshes β€” 85 MB

Download

pip install huggingface_hub
wget https://huggingface.co/datasets/tzuyuanlin/YCBMultiTrack/raw/main/download.py

python download.py --split all    --out ./YCBMultiTrack   # meshes + both splits, 10.1 GB
python download.py --split real   --out ./YCBMultiTrack   # real only, 8.9 GB
python download.py --split sim    --out ./YCBMultiTrack   # synthetic only, 1.1 GB
python download.py --split models --out ./YCBMultiTrack   # YCB meshes only, 85 MB

The helper downloads, verifies the sha256 recorded in sequences.json, and extracts. Or do it by hand:

from huggingface_hub import hf_hub_download
p = hf_hub_download("tzuyuanlin/YCBMultiTrack", "YCBMultiTrack-Sim.zip", repo_type="dataset")
sha256  afbc33285b2e2cd8350db0dd44626e195af0051462c8e2bc14ece06ad8df332e  YCBMultiTrack-Real.zip
sha256  e6d619cf8175a37e2cc8f25272a9b33fc7fef19581bd439117b1463eb8f90f7f  YCBMultiTrack-Sim.zip
sha256  ef77921d3f5b5616dfb9fc58d0f31ae07ec924c5b50316b2bbb400779d4064f6  YCBMultiTrack-Models.zip

Layout

Each split archive extracts to one directory per sequence:

YCBMultiTrack-Real/<sequence>/
β”œβ”€β”€ rgb/000000.png ...                  uint8 RGB, 640x480
β”œβ”€β”€ depth/000000.png ...                uint16 PNG, millimetres
β”œβ”€β”€ normalized_depth/000000.png ...     8-bit depth, for visualization only
β”œβ”€β”€ cam_K.txt                           3x3 intrinsics
β”œβ”€β”€ masks/<object>/000000.png ...       binary instance mask
β”œβ”€β”€ annotated_poses/<object>/000000.txt 4x4 T_mesh2cam, row-major
β”œβ”€β”€ is_mask_visible/<object>/
β”‚   └── is_mask_visible.npy             uint8 [N], per-frame
└── is_obj_in_image_labels/<object>/
    └── is_obj_in_image.npy             uint8 [N], per-frame

Frame ids are consistent across rgb/, depth/, normalized_depth/, masks/ and annotated_poses/: masks/<object>/000123.png and annotated_poses/<object>/000123.txt belong to rgb/000123.png. Real frame ids start at 000000, Sim frame ids at 000001; entry i of the per-frame .npy labels is the i-th frame in sorted order. Sequences are named after the objects they contain, so the object set is readable from the directory name.

Conventions

  • Depth is uint16 millimetres; divide by 1000 for metres. 0 means no depth reading. In Real, treat values above ~10 m as invalid if you back-project the full frame.

  • Real poses come from motion capture and are provided for every frame, including frames where the object is occluded or out of frame.

  • Poses are T_mesh2cam: a 4Γ—4 row-major matrix mapping the object mesh frame into the camera frame, in metres. They are defined against the textured_simple.obj meshes shipped in YCBMultiTrack-Models.zip β€” a different mesh origin will not align.

  • is_obj_in_image is 0 when the object is out of the image. An object that is only partly cut off at the image border is usually still labeled 1.

  • is_mask_visible is 0 when the object is occluded β€” hidden behind another object or the hand β€” or out of frame.

  • A frame marked 0 can still show part of the object, so its mask may be non-empty. Score only frames where both labels are 1, as the paper does.

  • Intrinsics are constant within each split:

    Split fx fy cx cy
    Real 614.429 615.125 323.473 239.058
    Sim 380.0 380.0 320.0 240.0

sequences.json carries the same facts machine-readably: per-sequence frame counts, objects, intrinsics, and per-object counts of frames where each label is 0.

Object meshes

The meshes are included β€” YCBMultiTrack-Models.zip (85 MB), so nothing has to be sourced separately:

YCBMultiTrack-Models/
β”œβ”€β”€ ATTRIBUTION.md
└── <object>/
    β”œβ”€β”€ textured_simple.obj        the mesh poses are defined against, in metres
    β”œβ”€β”€ textured_simple.obj.mtl    (072a_toy_airplane: textured.mtl)
    β”œβ”€β”€ texture_map.png            4096x4096
    └── points.xyz                 sampled point cloud, where available

Ten objects: 002_master_chef_can, 004_sugar_box, 005_tomato_soup_can, 006_mustard_bottle, 008_pudding_box, 010_potted_meat_can, 019_pitcher_base, 021_bleach_cleanser, 037_scissors, 072a_toy_airplane.

These are byte-for-byte copies of the model set shipped with HO3D-v3 (HO3D_V3/models/<object>/), which is HO3D's packaging of the YCB Object and Model Set; textured_simple.obj is the decimated variant also distributed with YCB-Video. If you already have HO3D-v3, this archive is redundant β€” point your model root at HO3D_V3/models instead. The full-resolution textured.obj (~51 MB/object) is omitted for size.

Redistributed under CC BY 4.0 β€” see ATTRIBUTION.md in the archive for citations.

Sequences

Real β€” Intel RealSense D435i, OptiTrack ground truth

Single-object sequences exercise out-of-frame disappearance; multi-object sequences add mutual occlusion, with easy/hard denoting the amount of occlusion and motion.

Percentages are the share of frames in which an object is occluded β€” hidden behind another object or a hand, or carried out of the image β€” i.e. where is_mask_visible or is_obj_in_image is 0. Those frames are excluded from scoring.

Sequence Frames Size on disk Objects (% frames occluded)
005_tomato_soup_can 1,420 824 MB tomato_soup_can (14%)
006_mustard_bottle 1,816 1,040 MB mustard_bottle (29%)
008_pudding_box 1,316 748 MB pudding_box (26%)
010_potted_meat_can 1,419 815 MB potted_meat_can (19%)
021_bleach_cleanser 1,561 905 MB bleach_cleanser (10%)
005_tomato_soup_can_008_pudding_box_easy 1,181 678 MB tomato_soup_can (11%), pudding_box (7%)
005_tomato_soup_can_008_pudding_box_hard 1,587 918 MB tomato_soup_can (11%), pudding_box (15%)
006_mustard_bottle_010_potted_meat_can_easy 1,342 794 MB mustard_bottle (2%), potted_meat_can (1%)
006_mustard_bottle_010_potted_meat_can_hard 1,769 1,034 MB mustard_bottle (15%), potted_meat_can (14%)
006_mustard_bottle_010_potted_meat_can_005_tomato_soup_can 1,366 793 MB tomato_soup_can (8%), mustard_bottle (9%), potted_meat_can (13%)
021_bleach_cleanser_005_tomato_soup_can_008_pudding_box 1,537 923 MB tomato_soup_can (4%), pudding_box (3%), bleach_cleanser (<1%)

Sim β€” Isaac Sim

In the synthetic split, objects stay in frame and occlude each other only slightly, so every frame is labeled observable. The split isolates pose accuracy without the occlusion-recovery challenge.

The split was rendered with Point2Pose-SyntheticDataGenerator, our open-source generator built on Isaac Lab / Isaac Sim. It spawns YCB objects, moves them along configurable trajectories, and records RGB, depth, intrinsics and per-object camera-frame poses, so you can use it to render additional sequences.

Sequence Frames Size on disk Objects
env0_019_pitcher_base 501 209 MB pitcher_base
env0_037_scissors_002_master_chef_can 501 207 MB scissors, master_chef_can
env1_004_sugar_box 501 206 MB sugar_box
env2_072a_toy_airplane_005_tomato_soup_can 701 308 MB toy_airplane, tomato_soup_can
env3_010_potted_meat_can_004_sugar_box 501 208 MB potted_meat_can, sugar_box

Ground truth

Real. Object poses come from an OptiTrack motion-capture system with markers attached to each object. The poses are transferred into the camera frame through a per-sequence camera-to-world extrinsic and a per-object mesh-to-rigid-body transform, which are obtained via per sequence optimization.

Sim. Poses are exact by construction.

Evaluation protocol

The protocol used in the Point2Pose paper:

  • Score a frame only where the object is observable β€” is_obj_in_image & is_mask_visible.
  • Align predictions to ground truth at the first valid frame (methods are tracked from an arbitrary initial frame, not from a canonical pose).
  • Report ADD and ADD-S, and their AUC with a 0.1 m threshold, against the textured_simple.obj meshes.

Reference implementation: experiments/ycbinisaac/ in the Point2Pose repository. YCBInIsaacReader in point2pose/io/sources/dataset/datareader.py reads this layout directly:

from point2pose.io.sources.dataset.datareader import YCBInIsaacReader

reader = YCBInIsaacReader("YCBMultiTrack/YCBMultiTrack-Real/005_tomato_soup_can")
rgb = reader.get_color(0)              # HxWx3 uint8
depth = reader.get_depth(0)            # HxW float, metres
poses = reader.get_gt_poses(0)         # {object_name: 4x4 T_mesh2cam}

License

Released under CC BY 4.0. The bundled YCB object meshes carry their own attribution from the YCB Object and Model Set, also CC BY 4.0 β€” see ATTRIBUTION.md inside YCBMultiTrack-Models.zip.

Citation

@inproceedings{lin2026point2pose,
  title     = {Point2Pose: Occlusion-Recovering 6D Pose Tracking and 3D Reconstruction
               for Multiple Unknown Objects via 2D Point Trackers},
  author    = {Lin, Tzu-Yuan and Lee, Ho Jae and Doherty, Kevin and
               Lee, Yonghyeon and Kim, Sangbae},
  booktitle = {European Conference on Computer Vision (ECCV)},
  year      = {2026}
}
Downloads last month
567