ViewRecDB-100K

This release contains 103,758 annotated examples for 3D viewpoint recommendation. The Dataset Viewer includes all 100,000 training examples, 2,758 validation examples, and 1,000 test examples. Given a photograph with suboptimal composition, the task is to predict how to adjust the camera to recover the viewpoint of an expert photograph.

How the pairs are generated

Each pair starts with an expert photograph from the Unsplash Full Dataset. We outpaint the photograph, reconstruct the scene in 3D, and render it from a different camera pose. After image restoration and hierarchical filtering, the rendered view serves as the input and the expert photograph serves as the target. The 1,000 test examples form ViewRecEval, where experts verified that the target has better composition than the input.

Example

Input · generated view	Target · expert photograph

The annotation describes the adjustment from the input viewpoint to the target viewpoint.

Field	Meaning
change_orientation	Whether to switch between landscape and portrait orientation.
pose[0:3]	Three translation components.
pose[3:7]	Four rotation quaternion components.
pose[7:9]	Two field-of-view scaling values.

Files

The three splits use the directories train/, val/, and test/. Each photo directory is named after its Unsplash photo ID. It contains one or two generated views. Match each image with its annotation by filename, for example generated_0.jpeg and generated_0.json. Both views share the same target, original.jpeg.

Obtain the target photographs

The release includes the generated views and annotations. The original photographs must be obtained separately under the Unsplash Full Dataset’s redistribution terms. Request access through Unsplash, download its metadata, then run the following command from the ViewRec3D repository.

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

Run the command again with val and test as the final directory in --id_root to retrieve the targets for those splits.