MrNeRF (@janusch_patas)
2025-06-06 | โค๏ธ 131 | ๐ 14
Revisiting Depth Representations for Feed-Forward 3D Gaussian Splatting
โข We pinpoint an unexposed yet critical issue that leads to lower-quality 3D Gaussians predicted by feed-forward 3DGS models, rooted in the long-standing discontinuity issue of depth.
โข We introduce a novel training loss, PM-Loss, designed to improve 3D Gaussian quality by leveraging the geometry prior from pointmaps obtained from pre-trained 3D reconstruction models.
โข Extensive experiments on existing feed-forward 3DGS models across two large-scale datasets demonstrate the effectiveness of our PM-Loss in enhancing the quality of both 3D Gaussians and rendered novel views.
๋ฏธ๋์ด
