USP-Gaussian: Unifying Spike-based Image Reconstruction, Pose Correction and Gaussian Splatting

Input GS w/ joint Rec w/ joint GS w/o joint Rec w/o joint Recon-Net 3DGS Operation Flow Gradient Flow Visual Comparison of the 3DGS and the Recon-Net (w/ & w/o Joint Learning) 22.3dB 24.5dB PSNR 26.4dB 27.1dB Figure 1. Left. Illustration of our USP-Gaussian framework, where the spike-based image Reconstruction Network (Recon-Net), camera poses, and 3DGS are collaboratively optimized signified by . Mid. Visual ablation showcasing the performance of Recon-Net and 3DGS with and without (w/ & w/o) the joint optimization strategy, with the ablation table depicted in Tab. 4 and the input formulated in Eq. (6). Right. Training curve comparison for Recon-Net and 3DGS with and without joint optimization.