Multi-View Video Renderings from 53K Animated 3D Objects

A 2024 dataset contains multi-view video sequences rendered from 53,000 animated 3D objects. Created by yanqinJiang, it was used to train the Animate3D model for animating 3D assets with video diffusion. For each object, 16 views are rendered with randomized camera elevation and azimuth offsets.

Use Cases

• Train multi-view video diffusion models like Animate3D using sequences of 16 rendered views per 3D object.
• Develop novel-view synthesis models by learning from the structured azimuth and elevation (elv, azi_start) camera parameters.
• Benchmark 3D animation generation from static models using the provided video sequences of 53K objects.
• Analyze the impact of randomized camera parameters (elv between 0-30 degrees, azi_start offset ±11.25 degrees) on video rendering consistency.

Strengths

• Large scale with 53,000 distinct animated 3D objects as source material.
• Structured multi-view capture with 16 video views per object.
• Precise, documented camera parameterization including elevation (elv) and azimuth start (azi_start).

Limitations

• Dataset is synthetic, rendered from 3D models, which may not capture real-world video artifacts.
• Limited camera elevation range (0-30 degrees) restricts viewpoint diversity.
• Unknown video resolution, frame rate, and total data size.

Provenance

Source

Hugging Face dataset repository by yanqinJiang.

Collection Method

Rendered from a collection of 53K animated 3D objects with defined camera parameters.

Freshness

Last updated October 2024.