Research map / 3D vision / Depth and stereo
Through the Looking Glass: Neural 3D Reconstruction of Transparent Shapes
Check your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace the visual hull with the full reconstruction pipeline | transparent shape reconstruction · Synthetic, 10 views | CD($10^{-4}$): 5.14 -> 2.00 | Helped |
| replace the visual hull with the full reconstruction pipeline | transparent shape reconstruction · Synthetic, 5 views | CD($10^{-4}$): 31.7 -> 6.30 | Helped |
| replace the visual hull with the full reconstruction pipeline | transparent shape reconstruction · Synthetic, 20 views | CD($10^{-4}$): 2.23 -> 1.20 | Helped |
| replace the visual hull with the full reconstruction pipeline | transparent shape reconstruction · Real transparent objects, mean | CD($10^{-4}$): 4.97 -> 4.00 | Helped |
| add rendering-error map and total-reflection mask inputs (ablation) | normal reconstruction · Synthetic, 10 views | N^1 median ($^\circ$): 3.5 -> 3.5 | No clear effect |
| add the cost volume (ablation) | normal reconstruction · Synthetic, 10 views | N^1 median ($^\circ$): 3.5 -> 3.4 | Helped |
| optimize the encoder latent vector using rendering error (ablation) | normal reconstruction · Synthetic, 10 views | Render Err.($10^{-2}$): 4.4 -> 2.9 | Helped |
| use Chamfer distance loss instead of nearest-point L2 loss (ablation) | point-cloud reconstruction · Synthetic, 10 views | CD($10^{-4}$): 2.17 -> 2.00 | Helped |
| use Chamfer distance loss instead of view-dependent L2 loss (ablation) | point-cloud reconstruction · Synthetic, 10 views | CD($10^{-4}$): 2.15 -> 2.00 | Helped |
| select views by rendering error rather than nearest-view selection (ablation) | point-cloud feature mapping · Synthetic, 10 views | CD($10^{-4}$): 2.04 -> 2.00 | Helped |
| select views by rendering error rather than average feature fusion (ablation) | point-cloud feature mapping · Synthetic, 10 views | CD($10^{-4}$): 2.03 -> 2.00 | Helped |
| use average-pooling instead of max-pooling (ablation) | PointNet++ point-cloud refinement · Synthetic, 10 views | CD($10^{-4}$): 2.09 -> 2.00 | Helped |
Closest papers by meaning
- Differentiable Neural Surface Refinement for Modeling Transparent Objects (CVPR 2024 · similarity 0.89)
- UniSDF: Unifying Neural Representations for High-Fidelity 3D Reconstruction of Complex Scenes with Reflections (NeurIPS 2024 · similarity 0.88)
- From Transparent to Opaque: Rethinking Neural Implicit Surfaces with $\alpha$-NeuS (NeurIPS 2024 · similarity 0.88)
- Multistable Shape from Shading Emerges from Patch Diffusion (NeurIPS 2024 · similarity 0.87)
- Generative Multiview Relighting for 3D Reconstruction under Extreme Illumination Variation (CVPR 2025 · similarity 0.87)
- Pygmalion Effect in Vision: Image-to-Clay Translation for Reflective Geometry Reconstruction (NeurIPS 2026 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.