SpaRe Dataset

SpaRe extends the DTU dataset. We believe that a fair evaluation protocol requires a larger sample of objects, including a subset such that test views are not publicly released. To this end, we replicate a DTU-alike setup in a synthetic Blender environment.

Scenes

We place all objects on a white platform and place them in a black box. We manually collected $102$ assets from BlenderKit. All the collected models are high-quality 3D assets belonging to a range of categories, including toys, cars and houses miniatures, home appliances and equipment, technology items, tools, sports equipment, plants, decorations etc. A breakdown of categories seen in SpaRe is presented below: The scenes span a large range of objects sized between 8cm and 50cm. We include items varying in texture and specularity (eg. a plushy elephant, a glass lamp, or a shiny plastic children’s toy).

Examples

Paper

For more detail we refer you to our dataset paper [1].

[1] Nazarczuk, M., Tanay, T., Catley-Chandar, S., Shaw, R., Timofte, R., Pérez-Pellitero, E.: AIM 2024 Sparse Neural Rendering Challenge: Dataset and Benchmark. In: Proceedings of the European Conference on Computer Vision (ECCV) Workshops (2024)