EBH Dataset

Overview

Hands frequently exhibit severe motion blur because of their fast and highly articulated movement. This makes 3D hand recovery challenging: a blurry image can contain visual evidence from multiple time steps, while the exact hand configuration at a particular instant remains ambiguous.

EBH provides real blurry hand images together with corresponding event streams and 3D hand annotations, enabling research on temporally ambiguous hand motion under realistic capture conditions.

01 — REAL BLUR Captures real-world hand motion blur rather than relying only on synthetic blur generation.

02 — EVENT STREAM Provides temporally dense event information to support motion understanding under severe blur.

03 — 3D ANNOTATION Includes MANO parameters, hand meshes, and 3D keypoints for sequence-level recovery.

Visual Examples

The EBH dataset captures synchronized blurry RGB images, event-stream observations, and 3D hand annotations. The examples below illustrate both the captured visual evidence and the corresponding 3D hand reconstruction target.

Example blurry hand images and event-stream observations from the EBH dataset — Captured Observation **Real blurry hand images with event information**

Example 3D hand mesh annotations from the EBH dataset — 3D Annotation **Hand mesh and temporal motion sequence**

Dataset Contents

Blurry Images

Real blurry hand images captured from ten individuals, organized by capture direction and camera view.

blur_images.zip

Event Streams

Event data divided into 1 ms intervals, providing temporally dense observations of rapid hand movement.

events.zip

3D Annotations

3D annotations including MANO parameters, hand meshes, and 3D keypoints for sequence-level supervision.

annotations.zip

Directory Structure

Blurry Images

blur_images/
├── left_xxxxxx/
│   └── cam_00/
│       └── xxxxxx.png
└── right_xxxxxx/
    └── cam_00/
        └── xxxxxx.png

Event Streams

events/
└── left_xxxxxx/
    ├── xxx_01.png
    ├── xxx_02.png
    ├── ...
    └── xxx_11.png

3D Annotations

annotations/
├── mano_params/
├── meshes/
└── keypoints_3d/

Download

The full EBH dataset is available through the official project Google Drive folder.

Download EBH Dataset ↓

Paper

The EBH dataset and EBHNet were introduced in:

Joonkyu Park, Gyeongsik Moon, Weipeng Xu, Evan Kaseman, Takaaki Shiratori, and Kyoung Mu Lee. 3D Hand Sequence Recovery from Real Blurry Images and Event Stream. ECCV 2024.

Read Paper ↗