EgoPressure: A Dataset for Hand Pressure and Pose Estimation in Egocentric Vision

CVPR 2025 (Highlight)

Yiming Zhao^1*, Taein Kwon^1*, Paul Streli^1*, Marc Pollefeys^1,2, Christian Holz¹

¹ ETH Zürich ² Microsoft
^* Equal contribution

A dataset for hand pressure and pose estimation in egocentric vision. Participants perform touch gestures on a pressure-sensing touchpad (Sensel Morph) while being recorded by a head-mounted egocentric camera and seven static Azure Kinects — providing synchronized RGB-D from 8 views, fine-grained touch pressure for every contact, and MANO hand mesh annotations.

• 📄 Paper: https://arxiv.org/abs/2409.02224
• 🛠️ Toolkit (download / load / visualize): https://github.com/eth-siplab/EgoPressure
• 🌐 Project page: https://yiming-zhao.github.io/EgoPressure/

Composition

Participants	21 (pseudonymized p_001 … p_021)
Sequences	1,344 — 32 gestures × 2 hands per participant
Gestures	palm/finger/index presses (high, low, no-contact), pinches, pinch-zoom, grasps, pull/push, drawing, typing + a calibration routine
Views	8 synchronized RGB-D cameras (1 egocentric 1920×1080 + 7 static 2560×1440) at 30 Hz
Frames	≈630K per camera stream (≈5M camera-frames overall)
Modalities	RGB, depth, hand masks, raw touch-pressure grid, UV pressure maps, MANO hand meshes (pose, shape, per-vertex refinement), per-frame egocentric camera pose

Quick start (toolkit)

The official toolkit handles selective download, loading every modality, geometry, and visualization:

pip install "egopressure[all] @ git+https://github.com/eth-siplab/EgoPressure"

egopressure list
egopressure download --participants p_001 --cameras d --modalities rgb,depth,pressure,pose
egopressure video p_001 p_001_press_palm_low_x5_right --out seq.mp4

from egopressure import EgoPressure

ep = EgoPressure.from_hub(participants=["p_001"], cameras=["d"],
                          modalities=["rgb", "depth", "pressure", "pose"])
seq = ep.sequence("p_001", "p_001_press_palm_low_x5_right")
frame = seq.load_frame(60, depth=True)      # rgb, depth, force, annotation
frame.show(camera="d", overlays=["mesh", "skeleton", "pressure"])

The toolkit's data reference documents every field, unit, and convention.

Download without the toolkit

Everything is sharded by participant / sequence / camera / modality, so plain huggingface_hub works too:

from huggingface_hub import snapshot_download

snapshot_download("eth-siplab/EgoPressure", repo_type="dataset",
                  local_dir="egopressure_data",
                  allow_patterns=["configs/p_001/*",
                                  "data/p_001/*/cam-d.color.parquet",
                                  "data/p_001/*/pressure.parquet",
                                  "data/p_001/*/annotation.parquet"])

Layout

configs/<participant>/<sequence>.json         # camera calibration + metadata
configs/<participant>/<sequence>_k4a/         # per-camera factory calibrations
data/<participant>/<sequence>/
    cam-d.color.parquet   cam-1.color.parquet ... cam-7.color.parquet
    cam-d.depth.parquet   cam-1.depth.parquet ... cam-7.depth.parquet
    cam-d.mask.parquet    cam-1.mask.parquet  ... cam-7.mask.parquet
    pressure.parquet                                 # Sensel grid + UV pressure
    annotation.parquet                               # MANO pose/mesh + ego pose

Each Parquet shard holds one row per frame (frame column).

Modalities

Shard	Content
cam-*.color	RGB frames as original JPEG bytes (ego d: 1920×1080; static 1..7: 2560×1440; undistorted)
cam-*.depth	512×512 uint16 depth in millimetres (PNG bytes), all 8 cameras; depth-sensor frame, covering the camera's color view (+32 px guard band) — 0 = no measurement
cam-*.mask	hand segmentation masks (PNG bytes)
pressure	raw Sensel Morph grid (105×185 float32, flattened) + normalised 224×224 UV pressure map with [min, max] range
annotation	MANO vertices (778×3), joint_position (21×3), betas (10), full_pose (48), transl, normals, displacement, per-static-camera visible_vertices (7×778), per-frame egocentric camera pose (ego_R 3×3, ego_T 3)

Geometry essentials

• World origin on the touchpad surface (240 × 137.5 mm active area, x–y plane at z = 0); MANO vertices and the ego pose are in metres, static-camera ModelViewMatrix extrinsics in millimetres.
• Images are undistorted → plain pinhole projection with the config intrinsics.
• Depth maps are in each camera's depth-sensor frame (not registered to color) and cover the region co-visible with that camera's color image; configs/<p>/<seq>_k4a/ holds the factory calibrations — the toolkit's register_depth_to_color() produces pixel-aligned RGB-D.
• Raw sensor counts convert to Newtons via counts / 1736 (cell pitch 1.25 mm).

License & citation

Released under CC BY-NC-SA 4.0 (non-commercial, academic). MANO-derived annotations are additionally subject to the MANO license.

@InProceedings{Zhao_2025_CVPR,
  author    = {Zhao, Yiming and Kwon, Taein and Streli, Paul and Pollefeys, Marc and Holz, Christian},
  title     = {EgoPressure: A Dataset for Hand Pressure and Pose Estimation in Egocentric Vision},
  booktitle = {Proceedings of the Computer Vision and Pattern Recognition Conference (CVPR)},
  month     = {June},
  year      = {2025},
  pages     = {27727--27738}
}