ViewRecDB-100K

Dataset Summary

ViewRecDB-100K is a 3D viewpoint recommendation dataset for AI photography. It contains 100K training samples and 1K test samples. Each sample consists of a paired suboptimal and optimal image, along with the corresponding 3D viewpoint change annotation. Given a suboptimal image, the task is to predict the 3D viewpoint change toward the optimal image.

The dataset is automatically constructed from the Unsplash Full Dataset. The test samples form a dedicated benchmark, where each optimal image has been verified by experts to have better photographic composition than the corresponding suboptimal image.

Dataset Structure

Data Structure

The dataset is organized into train, val, and test splits. Each sample folder contains one optimal image and 1 or 2 generated suboptimal images with corresponding viewpoint change annotations.

ViewRecDB-100K/
├── train/
│   └── photo_id/                # photo_id in Unsplash Full Dataset 
│       ├── original.jpeg        # optimal image
│       ├── generated_0.jpeg     # suboptimal image
│       ├── generated_0.json     # viewpoint change annotation
│       ├── generated_1.jpeg     # optional suboptimal image
│       └── generated_1.json     # optional viewpoint change annotation
├── val
└── test

Data Instances

A sample from the training set is provided below:

Suboptimal Image	Optimal Image	Viewpoint Change
		{ "change_orientation": true, "pose": [ 0.06223759800195694, -0.004903144668787718, -0.06147797778248787, -0.007686913013458252, -0.1884094774723053, -0.017775092273950577, 0.981899619102478, 0.4500928404485889, 0.651864323785646 ] }

### Data Fields

Each data instance contains the following fields:

• suboptimal_image / optimal_image: The input suboptimal image. It is an RGB image with a total pixel budget of 1024 × 1024.
• viewpoint_change: The 3D viewpoint change annotation from the suboptimal image to the optimal image.
  • change_orientation: A boolean value indicating whether the image orientation should be changed, i.e., switching between landscape and portrait.
  • pose: A 9-dimensional viewpoint descriptor:
    • pose[0:3]: translation vector.
    • pose[3:7]: rotation quaternion.
    • pose[7:9]: FoV scaling factor s, which denotes the FoV shrinking ratio, equivalent to the focal length scaling ratio.

Uses

Due to the redistribution restrictions of the Unsplash Full Dataset, the original.jpeg files in ViewRecDB-100K are not directly included in this release. Users need to apply for and download the Unsplash Full Dataset from the official source.

• After obtaining the Unsplash metadata files, specify their path with --csv_root and set --id_root to the split directory where the original images should be downloaded.

  python -m script.download \
    --csv_root /path/to/unsplash_metadata \
    --id_root /path/to/ViewRecDB-100K/train | val | test \
    --csv_pattern "*.csv*"
  

  Arguments:

  • --csv_root: Path to the directory containing the Unsplash metadata files.
  • --id_root: Path to the dataset split directory, such as train, val, or test.
  • --csv_pattern: Filename pattern used to match the metadata files, such as *.csv* or *.tsv*.