Research map / Detection and segmentation / CNN features
Joint Learning of Intrinsic Images and Semantic Segmentation
Anil S. Baslamisli, Thomas T. Groenestege, Partha Das, Hoang-An Le, Sezer Karaoglu, Theo Gevers
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 ground-truth albedo instead of RGB as SegNet input (ablation) | SegNet · NED test (synthetic, 16 classes) | mIoU: 0.5217 -> 0.5810 | Improved |
| use RGB plus albedo input instead of RGB only (ablation) | SegNet · NED test (synthetic, 16 classes) | mIoU: 0.5217 -> 0.5577 | Improved |
| add albedo to RGB input versus albedo alone (ablation) | SegNet · NED test (synthetic, 16 classes) | mIoU: 0.5810 -> 0.5577 | Hurt |
| add ground-truth segmentation labels as input to intrinsic decomposition (ablation) | ShapeNet · NED test (synthetic, 16 classes) | albedo MSE: 0.0094 -> 0.0076 | Improved |
| add ground-truth segmentation labels as input (ablation) | ShapeNet · NED test (synthetic, 16 classes) | shading DSSIM: 0.1303 -> 0.1312 | No clear effect |
| jointly learn intrinsics with segmentation instead of SegNet | segmentation network · NED test (synthetic, 16 classes) | mIoU: 0.5217 -> 0.6332 | Improved |
| jointly learn segmentation with intrinsics | ShapeNet · NED test (synthetic, 16 classes) | albedo MSE: 0.0094 -> 0.0030 | Improved |
| joint network instead of single-task cascade | segmentation · NED test (synthetic, 16 classes) | mIoU: 0.5810 -> 0.6332 | Improved |
| raise intrinsic loss weight w from 1 to 2 (ablation) | Joint model · NED test (synthetic, 16 classes) | mIoU: 0.533 -> 0.633 | Improved |
| raise intrinsic loss weight w from 2 to 4 (ablation) | Joint model · NED test (synthetic, 16 classes) | mIoU: 0.633 -> 0.611 | Hurt |
| set intrinsic loss weight w to 0.5 instead of 2 (ablation) | Joint model · NED test (synthetic, 16 classes) | global pixel accuracy: 0.9302 -> 0.7038 | Hurt |
| raise gamma_SMSE from 0.95 to 0.99 (ablation) | ShapeNet intrinsic loss · NED test (synthetic, 16 classes) | MSE: 0.0083 -> 0.0043 | Improved |
Closest papers by meaning
- Revisiting Deep Intrinsic Image Decompositions (CVPR 2018 · similarity 0.89)
- Reflectance Adaptive Filtering Improves Intrinsic Image Estimation (CVPR 2017 · similarity 0.89)
- Unsupervised Learning for Intrinsic Image Decomposition From a Single Image (CVPR 2020 · similarity 0.89)
- Towards Open-Vocabulary Remote Sensing Image Semantic Segmentation (AAAI 2025 · similarity 0.85)
- SemStereo: Semantic-Constrained Stereo Matching Network for Remote Sensing (AAAI 2025 · similarity 0.84)
- ReasonX: MLLM-Guided Intrinsic Image Decomposition (CVPR 2026 · 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.
