MrNeRF (@janusch_patas)
2024-11-25 | โค๏ธ 83 | ๐ 8
Dynamics-Aware Gaussian Splatting Streaming Towards Fast On-the-Fly Training for 4D Reconstruction
Contributions: โข We propose a novel three-stage pipeline for 4D dynamic spatial reconstruction that supports on-the-fly training and per-frame streaming. Our method builds on causal inputs and eliminates the need for full-length multi-view videos, thereby enhancing the practicability.
โข Our approach seamlessly integrates the three stages to optimize the reconstruction quality. By selectively inheriting newly introduced Gaussians from the preceding frame, effectively distinguishing dynamic primitives to allocate optimization emphasis, and enhancing areas with weak reconstruction using gradient information and optimization errors, our method ensures high-fidelity dynamic spatial reconstruction.
โข Extensive experiments demonstrate the superiority of our method in multiple aspects, including a 20% improvement in online training speed, superior reconstruction quality, and real-time rendering capability.
๋ฏธ๋์ด
Tags
domain-vision-3d domain-rendering domain-ai-ml domain-simulation