Professional-grade multi-view video database
MultiCamVideo is the team's companion research-grade dataset, containing 13,600 dynamic scenes, each shot by 10 synchronized cameras from different angles. The dataset pays special attention to the diversity of optical parameters, providing videos for four focal lengths from 18mm to 50mm, with three aperture settings from f/2.4 to f/10, totaling 1TB of data. All videos are in 1280×1280 resolution with a frame rate of 15fps, and the camera external reference matrix for each frame is precisely labeled.
The value of this dataset lies in the following: firstly, its strict spatio-temporal synchronization (error < 1ms) provides a reliable benchmark for multi-view algorithms; secondly, the rich optical variants help to study the influence of lens characteristics on the generation effect; and ultimately, the standardized JSON format of the camera parameters can be directly used for downstream tasks such as 3D reconstruction. Tests show that the model trained with this dataset improves PSNR metrics by an average of 12.51 TP3T on the new viewpoint generation task.
This answer comes from the articleReCamMaster: Rendering Tool for Generating Multi-View Videos from a Single VideoThe
English 
简体中文 
日本語 
Deutsch 
Português do Brasil