Version 2 of the round wooden garden table with a vase of dried flowers and a green ball, cut out of its garden automatically - now with a solid black tray, a dense underside and less haze.
What's new since v1:
- Training region: instead of only the exact SAM 3 mask, the splat was trained on a generous region around the table (mask outline with gaps filled, plus a margin), so the dark tray and the area right around the table were always learned, even in frames where SAM 3 missed them.
- Shape-guided cut: the FlashSplat cut was used as a guide rather than the final answer, because it stripped the faint layers that make the dark, shiny tray and the wood grain look real. Kept: everything inside the table's measured circle from the top down to just above the ground (top, tray, underside, legs); above the top, only what touches the vase and flowers; at ground level, only what touches the feet and ball.
- Detached specks removed; centred and scaled for viewing.
Same pipeline as before: SAM 3 text prompts ("round wooden table", "wooden vase", "dried flower arrangement", "green soccer ball", "table legs"), automatic refill of dropped-out frames and a per-frame parts check - no manual masking. Trained with Nerfstudio's splatfacto-big (gsplat), 30,000 steps.
Credits:
- Photos and camera positions: "garden" scene of the Mip-NeRF 360 dataset - Barron, Mildenhall, Verbin, Srinivasan & Hedman, CVPR 2022.
- Segmentation: SAM 3 (Meta AI).
- Cut-out guide: Shen, Yang & Wang, "FlashSplat: 2D to 3D Gaussian Splatting Segmentation Solved Optimally", ECCV 2024.
- 3D Gaussian Splatting: Kerbl et al., SIGGRAPH 2023. Training: Nerfstudio + gsplat. Cameras supplied with the dataset (computed with COLMAP).
Built step by step with help from Claude (Anthropic).