This dataset is derived from EPIC-KITCHENS-100 and released under CC BY-NC 4.0. By requesting access you agree that: (1) you will use it for non-commercial research only; (2) you will cite EPIC-KITCHENS-100 and ObjectForesight; (3) you will not attempt to identify, contact, or infer private information about any individual depicted in the egocentric video; and (4) you will comply with the EPIC-KITCHENS terms (https://epic-kitchens.github.io).

ObjectForesight: EPIC-KITCHENS-100 3D Object Trajectories

📄 Paper (arXiv:2601.05237) · 🤖 Model: raivn/ObjectForesight-EPIC-DiT · 🛠️ Curation/extraction pipeline: RustinS/ObjectForesight-Data

Per-object 3D manipulation extractions from EPIC-KITCHENS-100, used to train ObjectForesight, a 3D object-centric dynamics model that predicts future 6-DoF object poses and trajectories from short egocentric video.

This is the raw extraction, not pre-windowed trajectories. Each clip holds per-object segmentation masks, a 3D mesh, camera + depth, and per-frame 6-DoF object pose. The included loader (objectforesight_data) windows and filters these into the H-step trajectory samples ObjectForesight trains on (the paper's "2M+" windows are produced at load time from this data).

From each EPIC-KITCHENS-100 action clip: EgoHOS hand-object detection → SAM 2 mask refinement → SpaTrackerV2 depth & 3D tracks → TRELLIS object mesh → FoundationPose 6-DoF object pose.

At a glance

29,006 clips · 34,286 object trajectories (29,499 train / 4,787 val) · 34 participants · 633 videos.
~0.84 TiB, packaged as per-video tar shards in data/shards/PXX_YY.tar (each tar mirrors the relative PXX_YY_Z/… tree).
data/metadata/trajectories.parquet indexes every trajectory (ids, split, n_frames, extrinsics convention, shard).

Per-trajectory artifacts (inside each shard)

path	description
`PXX_YY_Z/action.mp4`	RGB action clip (~960×544), from EPIC-KITCHENS-100
`PXX_YY_Z/spatracker.npz`	SpaTrackerV2: `intrinsics`, `extrinsics`, `depths` (see note), sparse `coords`/`visibs` tracks
`PXX_YY_Z/egohos/*`	EgoHOS hand & object masks
`…/objects/<obj>/masks.npz`, `clean_masks.npz`, `vas_masks.npz`, `vas_clean_masks.npz`	SAM 2 / video-amodal object masks
`…/objects/<obj>/clean_cropped_frames.npz`	cleaned object crops (TRELLIS input)
`…/objects/<obj>/trellis/model.glb`	TRELLIS 3D object mesh
`…/objects/<obj>/foundationpose10/poses.npz`	per-frame 6-DoF pose: `frame_ids (T,)`, `T_c_o (T,4,4)`, `init_from_frame`
`…/objects/<obj>/foundationpose10/{run_summary.json, track_log.csv}`	pose metadata + per-frame quality metrics
`…/objects/<obj>/moved_by_hand.txt`	hand-contact flag

Camera & pose conventions

data/metadata/extrinsics_conv.json tags each clip c2w or w2c; respect it per clip (most are c2w; 782 trajectories are w2c).
poses.npz T_c_o is the 4×4 object-in-camera transform per tracked frame; frame_ids are the (possibly non-contiguous) source frame indices.

Depth precision note

The depths in spatracker.npz are stored as float16 to keep the release around half the size. EPIC manipulation happens at close range, under 2.4 m, where float16 rounding stays on the order of 1 mm. That is finer than SpaTrackerV2's depth is actually accurate to, so nothing useful is lost, and np.load reads the arrays back as normal. If you need bit-exact float32 depth, re-run SpaTrackerV2 on the source EPIC clips.

Usage

huggingface-cli download raivn/ObjectForesight-EPIC --repo-type dataset --local-dir of-epic
cd of-epic
python examples/prepare.py            # untar shards into a flat ./manip_data tree
pip install -e .                      # the windowing/filtering loader (numpy, torch, opencv-python, omegaconf, lz4, rich, tqdm)
python examples/load_trajectories.py  # builds H-step trajectory windows

from objectforesight_data.data.datasets.dataset_epic import SceneSequenceDataset
ds = SceneSequenceDataset(dataset_root="manip_data", H=8)  # windows + filters the raw clips
sample = ds[0]   # one H-step object-trajectory window (scene point cloud, poses, target, mesh path, …)

Splits

data/metadata/splits/{train,val}.txt lists object-level keys (PXX_YY_Z/objects/<obj>). 29,499 train / 4,787 val.

License & attribution

Released under CC BY-NC 4.0, inherited from EPIC-KITCHENS-100. Non-commercial research use only. This is a derivative work (hand-object segmentation, 3D mesh reconstruction, depth, and 6-DoF pose estimation applied to EPIC-KITCHENS-100 frames). You must give appropriate credit and comply with the EPIC-KITCHENS terms.

Citation

@article{soraki2026objectforesight,
  title   = {ObjectForesight: Predicting Future 3D Object Trajectories from Human Videos},
  author  = {Soraki, Rustin and Bharadhwaj, Homanga and Farhadi, Ali and Mottaghi, Roozbeh},
  journal = {arXiv preprint arXiv:2601.05237},
  year    = {2026}
}
@article{damen2022rescaling,
  title   = {Rescaling Egocentric Vision: Collection, Pipeline and Challenges for EPIC-KITCHENS-100},
  author  = {Damen, Dima and Doughty, Hazel and Farinella, Giovanni Maria and Furnari, Antonino and
             Kazakos, Evangelos and Ma, Jian and Moltisanti, Davide and Munro, Jonathan and
             Perrett, Toby and Price, Will and Wray, Michael},
  journal = {International Journal of Computer Vision (IJCV)},
  year    = {2022}
}

Built with (please also cite): EgoHOS · SAM 2 · TRELLIS · SpaTrackerV2 · FoundationPose. See the curation/extraction pipeline at RustinS/ObjectForesight-Data for exact references and per-step details.

Downloads last month: 23

Total file size:

942 GB

Models trained or fine-tuned on raivn/ObjectForesight-EPIC

raivn/ObjectForesight-EPIC-DiT

0.2B • Updated about 21 hours ago • 20 • 1

Paper for raivn/ObjectForesight-EPIC

ObjectForesight: Predicting Future 3D Object Trajectories from Human Videos

Paper • 2601.05237 • Published Mar 22