README.md

metadata

license: cc-by-4.0
dataset_info:
  features:
    - name: image
      dtype: image
    - name: scene
      dtype: string
    - name: category
      dtype: string
    - name: image_name
      dtype: string
    - name: camera_model
      dtype: string
    - name: width
      dtype: int64
    - name: height
      dtype: int64
    - name: fx
      dtype: float64
    - name: fy
      dtype: float64
    - name: cx
      dtype: float64
    - name: cy
      dtype: float64
    - name: qw
      dtype: float64
    - name: qx
      dtype: float64
    - name: qy
      dtype: float64
    - name: qz
      dtype: float64
    - name: tx
      dtype: float64
    - name: ty
      dtype: float64
    - name: tz
      dtype: float64
  splits:
    - name: evaluation
      num_bytes: 1093853358
      num_examples: 335
  download_size: 1093903288
  dataset_size: 1093853358
configs:
  - config_name: default
    data_files:
      - split: evaluation
        path: data/evaluation-*

The D2P dataset

The D2P dataset is a dataset based on the Depth2Pose monocular depth estimation benchmark, a pose-based evaluation of MDEs without ground-truth depth. The dataset contains challenging scenes beyond the distribution of common training data, together with a simple and extensible evaluation framework, presented on the github page. The scenes are divided into two categories: statues and vegetation. Undistorted images and reconstructions in standard colmap format is provided for each scene, together with a list of image pairs used for the evaluation.

This D2P dataset example contains a small version of the original D2P dataset intended for easier overview. Here, only a subset of the scenes are included. The structure within the scenes is the same. To use the D2P Dataset, please, visit the page of the original dataset.

paper (coming later) | github | webpage

Dataset Structure

d2p_dataset_example                           
├── statues/                           
│   ├── scene1/                          
│   │   ├── images/                       
│   │   │   ├── img1.png
│   │   │   ├── img2.png   
│   │   │   └── ... 
│   │   ├── sparse/
│   │   │   ├── cameras.txt
│   │   │   ├── frames.txt
│   │   │   ├── images.txt
│   │   │   ├── points3D.txt
│   │   │   └── rigs.txt
│   │   ├── scene1_image_list.txt
│   │   └── scene1_image_pairs.txt
│   ├── scene2/ 
│   │   └── ... 
│   └── ... 
└──  vegetation/

Dataset Fields

Each scene contains:

• images/: RGB images
• sparse/: COLMAP reconstruction files:
  • camera parameters
  • frames
  • image poses
  • 3D points
  • rigs
• scene1_image_list.txt: List of all images used for the benchmark, found in the images/ folder
• scene1_image_pairs.txt: List of all image pairs used for the benchmark, for which realtive pose is evaluated

Direct Use

Benchmarking monocular depth estimators. For the current leaderboard, see the Depth2Pose webpage

Load with 🤗 Datasets

from datasets import load_dataset

ds = load_dataset("floodgab/d2p_dataset_example")
print(ds["validation"][0])

Loading Example

To download the Depth2Pose dataset

from huggingface_hub import snapshot_download

path = snapshot_download("floodgab/d2p_dataset_example")

Citation

If you use Depth2Pose in your research or find our work helpful, please cite

@misc{depth2pose,
  title={{Depth2Pose}: A Pose-Based Benchmark for Monocular Depth Estimation without Ground-Truth Depth},
  author={Kocur, Viktor and Aung, Sithu and Flood, Gabrielle and Ding, Yaqing and Bujnak, Lukas and Sattler, Torsten and Kukelova, Zuzana},
  year={2026},
}