720.20 c
Description: "Triplane Meets Gaussian Splatting" introduces an efficient 3D reconstruction method from single images using transformer-based networks for rapid feed-forward inference. The approach combines a point decoder and a triplane decoder, leveraging a hybrid Triplane-Gaussian representation for fast rendering and high-quality results. This hybrid design outperforms existing methods in speed and quality by balancing between implicit and explicit 3D representatio.
#700#Research_Paper#720#Computer_Vision#720.20#3D_Reconstruction#720.20 c#Triplane_Meets_Gaussian_Splatting:_Fast_and_Generalizable_Single-View_3D_Reconstruction_with_Transformers
Output: 3D
Type: .splat
Input: Image
Code: https://github.com/VAST-AI-Research/TriplaneGaussian
Demo: https://huggingface.co/spaces/VAST-AI/TriplaneGaussian
논문: https://arxiv.org/pdf/2312.09147
매우빠름 A10G 기준 한 30초 되는듯.
근데 역시 이미지의 안 보이는 면이 아쉬움
근데 화질이나 질감 빛차이가 덜할수록 오브젝트로 만들어질때는 쉬운듯? 테스트필요
원본 이미지
생성된 오브젝트 앞 뒤
원본이미지
생성한 오브젝트