Research map / Detection and segmentation / Open-set and few-shot detection
A 4D Light-Field Dataset and CNN Architectures for Material Recognition
Ting-Chun Wang, Jun-Yan Zhu, Ebi Hiroaki, Manmohan Chandraker, Alexei A. Efros, Ravi Ramamoorthi
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 angular filter light-field input instead of single 2D view | MINC VGG-16 · Light-field material dataset (128x128 patches) | accuracy (%): 70.2 -> 77.0 | Improved |
| add view-pooling of per-view features instead of single 2D view | MINC VGG-16 · Light-field material dataset (128x128 patches) | accuracy (%): 70.2 -> 70.0 | Hurt |
| average predictions across views instead of one view | MINC VGG-16 · Light-field material dataset (128x128 patches) | accuracy (%): 70.2 -> 70.5 | No clear effect |
| stack all views across RGB channels | MINC VGG-16 · Light-field material dataset (128x128 patches) | accuracy (%): 70.2 -> 72.8 | Improved |
| use EPI input | MINC VGG-16 · Light-field material dataset (128x128 patches) | accuracy (%): 70.2 -> 72.3 | Improved |
| apply angular filter at input rather than after layer 1 (ablation) | Angular filter · Light-field material dataset (128x128 patches) | accuracy (%): 76.6 -> 77.8 | Improved |
| apply angular filter at input rather than after layer 2 (ablation) | Angular filter · Light-field material dataset (128x128 patches) | accuracy (%): 73.7 -> 77.8 | Improved |
| use single-layer angular filter instead of stacked 3x3 filters (ablation) | Angular filter · Light-field material dataset (128x128 patches) | accuracy (%): 74.8 -> 77.8 | Improved |
| use 64 angular output channels instead of 3 (ablation) | Angular filter · Light-field material dataset (128x128 patches) | accuracy (%): 71.6 -> 77.8 | Improved |
| use 64 angular output channels instead of 147 (no compression) (ablation) | Angular filter · Light-field material dataset (128x128 patches) | accuracy (%): 72.8 -> 77.8 | Improved |
| use light-field input at equal pixel count (downsampled) versus 2D | Light-field CNN · Light-field patches, downsampled to equal pixel count | accuracy (%): 70.7 -> 75.2 | Improved |
| use light-field input with fine-tuning on BTF data versus 2D | Light-field CNN · BTF database (fine-tuned) | accuracy (%): 67.7 -> 73.0 | Improved |
Closest papers by meaning
- Deep Material Recognition in Light-Fields via Disentanglement of Spatial and Angular Information (ECCV 2020 · similarity 0.90)
- Harnessing the Power of Foundation Models for Accurate Material Classification (CVPR 2026 · similarity 0.87)
- Hierarchical Material Recognition from Local Appearance (ICCV 2025 · similarity 0.87)
- OpticalNet: An Optical Imaging Dataset and Benchmark Beyond the Diffraction Limit (CVPR 2025 · similarity 0.86)
- Neural LightRig: Unlocking Accurate Object Normal and Material Estimation with Multi-Light Diffusion (CVPR 2025 · similarity 0.85)
- Practical Measurements of Translucent Materials with Inter-Pixel Translucency Prior (CVPR 2024 · similarity 0.85)
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.
