Research map / Image generation / Image restoration
Single Image Intrinsic Decomposition without a Single Intrinsic Image
Wei-Chiu Ma, Hang Chu, Bolei Zhou, Raquel Urtasun, Antonio Torralba
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 |
|---|---|---|---|
| add disentangled reconstruction loss Ldis (ablation) | unsupervised intrinsic net (Lrec only) · MIT Intrinsics | albedo LMSE: 0.0158 -> 0.0141 | Improved |
| add disentangled reconstruction loss Ldis (ablation) | unsupervised intrinsic net (Lrec only) · MIT Intrinsics | albedo MSE: 0.0362 -> 0.0346 | Improved |
| add Retinex loss (ablation) | Lrec + Ldis model · MIT Intrinsics | albedo LMSE: 0.0141 -> 0.0116 | Improved |
| add Retinex loss (ablation) | Lrec + Ldis model · MIT Intrinsics | shading LMSE: 0.0098 -> 0.0095 | Improved |
| add Retinex loss (ablation) | Lrec + Ldis model · MIT Intrinsics | albedo MSE: 0.0346 -> 0.0313 | Improved |
| fine-tune unsupervised model with ground truth (100%) | Our method · ShapeNet | average LMSE: 0.0060 -> 0.0020 | Improved |
| fine-tune unsupervised model with ground truth (100%) | Our method · MIT Intrinsics | average LMSE: 0.0105 -> 0.0063 | Improved |
| fine-tune unsupervised model with ground truth (100%) | Our method · MIT Intrinsics | average MSE: 0.0260 -> 0.0130 | Improved |
| fine-tune unsupervised model with ground truth | Our method · MIT Intrinsics (illumination invariance) | MPRE (x10^-4): 3.2341 -> 2.4151 | Improved |
| use learned soft assignment mask instead of traditional Retinex for gradient classification | Our method · MIT Intrinsics | gradient pseudo-GT error: 2.6 x 10-4 -> 1.7 x 10-4 | Improved |
| use only 50% of ground-truth supervision during fine-tuning | Our method · ShapeNet | MSE/LMSE: state-of-the-art with only 50% of ground truth | Improved |
Closest papers by meaning
- Revisiting Deep Intrinsic Image Decompositions (CVPR 2018 · similarity 0.92)
- Unsupervised Intrinsic Image Decomposition Using Internal Self-Similarity Cues (TPAMI 2021 · similarity 0.90)
- Self-Supervised Intrinsic Image Decomposition (NeurIPS 2017 · similarity 0.89)
- Learning Intrinsic Image Decomposition From Watching the World (CVPR 2018 · similarity 0.89)
- Single Image Reflection Separation via Dual-Stream Interactive Transformers (NeurIPS 2024 · similarity 0.85)
- Intrinsic Image Decomposition for Robust Self-supervised Monocular Depth Estimation on Reflective Surfaces (AAAI 2025 · 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.
