MrNeRF (@janusch_patas)
2025-12-16 | โค๏ธ 65 | ๐ 14
Fast 2DGS: Efficient Image Representation with Deep Gaussian Prior
Contributions: โข We propose Deep Gaussian Prior, an initialization strategy learned through an iterative optimization-sampling loop. By simulating the optimization trajectory, our method captures a content-aware distribution that breaks the uniform bias of random initialization, significantly accelerating convergence.
โข We present a streamlined framework that anchors Gaussian cardinality (K) to compression rates. By employing a lightweight backbone without complex feature engineering, we transform the ill-posed Gaussian allocation into a tractable, batch-parallelizable learning task.
โข We demonstrate that our framework achieves superior trade-offs between reconstruction quality, inference latency, and cross-dataset generalizability compared to existing Gaussian image approaches.
๋ฏธ๋์ด
