SIGGRAPH Asia 2025｜30FPS普通相机恢复200FPS细节，4D重建方案来了

Core Viewpoint - The article discusses advancements in 4D reconstruction technology, specifically focusing on a new method that combines asynchronous capture with a video diffusion model to enhance the quality of high-speed dynamic scene reconstruction using low-cost hardware [3][10]. Group 1: Hardware Innovation - The asynchronous capture method allows multiple cameras to work in a "relay" fashion, overcoming the speed limitations of individual cameras. This method introduces a slight delay in the activation of different cameras, effectively doubling the frame rate from 25 FPS to 100 FPS or even reaching 200 FPS by organizing the cameras into groups [5][6][8]. Group 2: Software Innovation - A video diffusion model is employed to address the "sparse view" problem that arises from asynchronous capture, which can lead to visual artifacts in the initial 4D reconstruction. This model is trained to repair these artifacts and enhance video quality by utilizing the spatio-temporal context provided by the input video [9][10][13]. Group 3: Overall Process - The method integrates hardware capture with AI algorithms in an iterative optimization framework. The process includes initial reconstruction using asynchronous capture, generating pseudo ground truth videos, enhancing these videos with the diffusion model, and optimizing the 4D Gaussian model based on the enhanced output [14][15][17]. Group 4: Method Effectiveness - The proposed method outperforms several state-of-the-art techniques in key metrics such as Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index (SSIM), and Learned Perceptual Image Patch Similarity (LPIPS) across two public datasets, demonstrating its effectiveness in producing high-quality 4D reconstructions [19][21]. Group 5: Real-World Validation - A multi-view capture system consisting of 12 cameras operating at 25 FPS was established to validate the method in real-world scenarios. The experiments confirmed that the approach could robustly reconstruct high-quality, temporally consistent 4D content even in complex asynchronous capture environments [22].