Perceptual-Driven SuGaR: Real-Time High-Fidelity 3D Reconstruction via Adaptive Gaussian Splatting

This study addresses the geometric precision deficiency in the SuGaR algorithm, which sacrifices accuracy for mesh editability. By integrating a multi-scale perceptual loss mechanism and adversarial training framework into the loss function, coupled with a curriculum learning strategy, we achieve a significant improvement in geometric fidelity. Experimental results demonstrate a 2.75 dB enhancement in PSNR and 0.0024 increase in SSIM compared to the native SuGaR implementation.