caiyuanhao1998/HDR-GS
"HDR-GS: Efficient High Dynamic Range Novel View Synthesis at 1000x Speed via Gaussian Splatting" (NeurIPS 2024)
What it solves
This project tackles the problem of generating new viewpoints of a scene from existing images, but for high dynamic range (HDR) content. Traditional methods like NeRF are slow, often taking minutes per view. HDR-GS instead uses Gaussian splatting, a faster technique, to create HDR novel views at a claimed 1000x speedup. It also provides recalibrated camera poses and datasets, since existing HDR datasets were not compatible with the 3D Gaussian Splatting pipeline.
How it works
HDR-GS represents a 3D scene as a collection of Gaussian primitives, each with parameters like position, color, and opacity. For HDR, the method learns a tone mapper to convert high dynamic range radiance values to standard images. The pipeline includes structure-from-motion to recalibrate camera poses and generate initial point clouds. The code provides training scripts for both synthetic and real scenes, and testing scripts that load pretrained models to render novel views or spiral videos.
Who it’s for
This is aimed at researchers and developers working on novel view synthesis, HDR imaging, or real-time rendering. It is also useful for anyone wanting to experiment with Gaussian splatting or reproduce the results from the NeurIPS 2024 paper. Users should be comfortable with command-line tools and have a GPU for training.
Highlights
- Accepted at NeurIPS 2024
- Achieves 1000x speedup over previous NeRF-based HDR methods
- Includes code for training and testing on both synthetic and real scenes
- Provides recalibrated datasets and pretrained models via Google Drive, Baidu Disk, and Hugging Face
- Offers interactive comparison results on the project page
- Supports direct model loading and spiral video rendering with the provided scripts
Related
- Project
- Project
- Project
- Project