Research map / 3D vision / Depth and stereo
SeeGroup: Multi-Layer Depth Estimation of Transparent Surfaces via Self-Determined Grouping
Hongyu Wen, Jia Deng
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 |
|---|---|---|---|
| use SeeGroup (RD + max-mixture intensity + permutation-invariant loss) over Multi-head head | Depth Anything V2 multi-layer depth · LayeredDepth test (All) | quadruplet accuracy: 61.34 -> 70.09 | Improved |
| add recurrent decomposition instead of GRU (ablation) | SeeGroup · LayeredDepth val (ViT-L) | quadruplet accuracy All: 67.82 -> 71.50 | Improved |
| use recurrent decomposition instead of independent multi-head on deep layers (ablation) | SeeGroup · LayeredDepth val Layer 5 | quadruplet accuracy: 39.64 -> 66.14 | Improved |
| use recurrent decomposition instead of multi-head overall (ablation) | SeeGroup · LayeredDepth val (All) | quadruplet accuracy: 73.04 -> 71.50 | Mixed |
| use max-mixture instead of weighted mixture (ablation) | SeeGroup (ViT-S) · LayeredDepth val (All) | quadruplet accuracy: 56.12 -> 69.03 | Improved |
| use max-mixture (permutation invariant) instead of depth-ordered sorted mixture (ablation) | SeeGroup (ViT-S) · LayeredDepth val (All) | quadruplet accuracy: 52.68 -> 69.03 | Improved |
| add coverage loss to intensity and GM losses (ablation) | SeeGroup (ViT-S) · LayeredDepth val (All) | quadruplet accuracy: 68.36 -> 69.03 | Improved |
| add gradient matching loss to L1 (ablation) | L1 loss · LayeredDepth val (All) | quadruplet accuracy: 42.49 -> 57.00 | Improved |
| use intensity loss instead of L1 with GM (ablation) | SeeGroup (ViT-S) · LayeredDepth val (All) | quadruplet accuracy: 57.00 -> 68.36 | Improved |
| use full loss instead of SiLog (ablation) | SeeGroup (ViT-S) · LayeredDepth val (All) | quadruplet accuracy: 48.20 -> 69.03 | Improved |
| use SeeGroup instead of Multi-head DA v2 on synthetic in-domain data | SeeGroup · LayeredDepth-Syn val Layer 1 | AbsRel: 8.24 -> 9.82 | Hurt |
| use RD instead of GRU for model size (ablation) | RD · model size | #Param: 731M -> 356M | Improved |
Closest papers by meaning
- Seeing and Seeing Through the Glass: Real and Synthetic Data for Multi-Layer Depth Estimation (ICCV 2025 · similarity 0.91)
- SeeClear: Reliable Transparent Object Depth Estimation via Generative Opacification (ECCV 2026 · similarity 0.88)
- Learning Depth Estimation for Transparent and Mirror Surfaces (ICCV 2023 · similarity 0.86)
- 3DReflecNet: A Large-Scale Dataset for 3D Reconstruction of Reflective, Transparent, and Low-Texture Objects (CVPR 2026 · similarity 0.86)
- TSPE-GS: Probabilistic Depth Extraction for Semi-Transparent Surface Reconstruction via 3D Gaussian Splatting (AAAI 2026 · similarity 0.85)
- From Transparent to Opaque: Rethinking Neural Implicit Surfaces with $\alpha$-NeuS (NeurIPS 2024 · similarity 0.84)
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.
