EgoXtreme: A Dataset for Robust Object Pose Estimation in Egocentric Views under Extreme Conditions

📖 Dataset Information

EgoXtreme is a novel large-scale dataset designed for robust egocentric 6D object pose estimation under extreme environmental conditions. The dataset comprises approximately 1.3 million frames with a total duration of 775.5 minutes (~12.9 hours). It was captured at 30 fps using Aria glasses, providing high-resolution 1408 x 1408 raw fisheye RGB images along with their undistorted versions.

The dataset features 15 participants performing diverse interactions with 13 different objects (including sports equipment, assembly blocks, and emergency supplies). It is divided into training (518.8 min), validation (80.7 min), and test (176 min) sets across three challenging scenarios: Industrial Maintenance, Sports, and Emergency Rescue.

Note on Test Set: For fair evaluation, the GT annotations for the test set are withheld. The test images can be downloaded from our separate repository: taegyoun88/egoxtreme-test.

🛠️ Sample Usage

The official repository provides tools to process and visualize the data.

Undistortion

Due to the large file size, undistorted versions of the data are generated via scripts. To generate undistorted RGB images and masks:

# Process a specific scene
python tools/undistortion.py --data_dir ./data/train --scene_id 000000

# Process all scenes in train/test set
python tools/undistortion.py --data_dir ./data/train --all

Visualization

To visualize the Ground Truth 6D pose on the images:

# Visualize specific scene (Add --undist for undistorted images, --im_id for single frame)
python tools/visualization.py --data_dir ./data/test --scene_id 000000 --models_dir ./models [--undist] [--im_id 0]

🎛️ Scenario Configurations

The detailed configurations of illumination and environmental conditions for each scenario are summarized below:

Scenario	Standard (normal, middle, high)	Extreme (low)	Extreme (head)	Extreme (flash)	Extreme (warning)	Extreme (green)	Smoke	Object
Maintenance	✔️	✔️	✔️	✔️			✔️	5
Sports	✔️	✔️					✔️	5
Emergency	✔️	✔️			✔️	✔️	✔️	3

Below is the mapping of Scene IDs to their corresponding scenarios across the dataset splits:

Split	Scenario	Scene IDs
Train	Maintenance	000000 - 000211
	Sports	000212 - 000417
	Emergency	000418 - 000573
Validation	Maintenance	000000 - 000039
	Sports	000040 - 000067
	Emergency	000068 - 000079

📁 Dataset Structure & Format

All files (*.json) and 3d model information follow the BOP format.

The structure of the data hosted here is organized as follows:

EgoXtreme
├── models/                                      # 3D CAD models (.ply) and info
└── data/
    ├── train/
    │   ├── 000000/                              # Scene ID
    │   │   ├── rgb/                             # Raw fisheye RGB images
    │   │   ├── mask/                            # Full object masks
    │   │   ├── scene_camera.json
    │   │   ├── scene_gt.json
    │   │   ├── scene_gt_info.json
    │   │   └── scene_camera_undist.json
    │   └── ...
    └── val/ ...
        ├── 000000/
        └── ...

Citation

@inproceedings{egoxtreme2026,
  title={EgoXtreme: A Dataset for Robust Object Pose Estimation in Egocentric Views under Extreme Conditions},
  author={Yoon, Taegyoon and Han, Yegyu and Ji, Seojin and Park, Jaewoo and Kim, Sojeong and Kwon, Taein and Kim, Hyung-Sin},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  year={2026}
}