Research map / 3D vision / Depth and stereo
Learning to See Through Obstructions With Layered Decomposition
Yu-Lun Liu, Wei‐Sheng Lai, Ming–Hsuan Yang, Yung‐Yu Chuang, Jia‐Bin Huang
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 |
|---|---|---|---|
| predict uniform flow fields as initialization instead of zero (ablation) | flow decomposition · Vimeo-90k validation | validation loss: 0.478 -> 0.197 | Improved |
| predict uniform flow fields instead of dense flow fields (ablation) | flow decomposition · Vimeo-90k validation | validation loss: 0.236 -> 0.197 | Improved |
| use image reconstruction network instead of temporal mean filtering (ablation) | fusion · Vimeo-90k validation | validation loss: 0.652 -> 0.197 | Improved |
| use image reconstruction network instead of temporal median filtering (ablation) | fusion · Vimeo-90k validation | validation loss: 0.555 -> 0.197 | Improved |
| add pre-training with synthetic data (ablation) | online-optimization-only model · Vimeo-90k validation | validation loss: 0.468 -> 0.197 | Improved |
| add online optimization on real data (ablation) | pre-trained model · Vimeo-90k validation | validation loss: 0.283 -> 0.197 | Improved |
| use coarse-to-fine with 5 pyramid levels instead of 1 (ablation) | pyramid levels · Vimeo-90k validation | validation loss: 0.382 -> 0.283 | Improved |
| increase pyramid levels from 5 to 6 (ablation) | pyramid levels · Vimeo-90k validation | validation loss: 0.283 -> 0.281 | No clear effect |
| add the registered frame input feature (ablation) | input features · Vimeo-90k validation | validation loss: 0.416 -> 0.283 | Improved |
| add the difference map input feature (ablation) | input features · Vimeo-90k validation | validation loss: 0.288 -> 0.283 | Improved |
| use proposed realistic training data generation (ablation) | training data · Controlled sequences Stone and Toy | NCC (background): 0.9390 -> 0.9428 | Improved |
| use proposed realistic training data generation (ablation) | training data · Controlled sequences Stone and Toy | NCC (reflection): 0.5886 -> 0.6032 | Improved |
Closest papers by meaning
- Learning to See Through Obstructions (CVPR 2020 · similarity 0.99)
- Instruct2See: Learning to Remove Any Obstructions Across Distributions (ICML 2025 · similarity 0.88)
- A Generic Deep Architecture for Single Image Reflection Removal and Image Smoothing (ICCV 2017 · similarity 0.87)
- Seeing Deeply and Bidirectionally: A Deep Learning Approach for Single Image Reflection Removal (ECCV 2018 · similarity 0.86)
- Neural Spline Fields for Burst Image Fusion and Layer Separation (CVPR 2024 · similarity 0.86)
- Gated Fields: Learning Scene Reconstruction from Gated Videos (CVPR 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.
