Research map / Image generation / Image restoration
Indices Matter: Learning to Index for Deep Image Matting
Hao Lu, Yutong Dai, Chunhua Shen, Songcen Xu
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 max-pooling indices (unpooling) instead of bilinear upsampling (ablation) | SegNet VGG16 · Composition-1k | SAD: 122.4 -> 54.6 | Improved |
| add max-pooling indices instead of bilinear upsampling (ablation) | SegNet MobileNetv2 · Composition-1k | SAD: 78.6 -> 60.7 | Improved |
| add skip fusion of low-level features (ablation) | SegNet MobileNetv2 with indices · Composition-1k | SAD: 60.7 -> 57.1 | Improved |
| use concatenation instead of skip fusion (ablation) | UNet MobileNetv2 with indices · Composition-1k | SAD: 57.1 -> 54.7 | Improved |
| add ASPP context (ablation) | SegNet MobileNetv2 with indices · Composition-1k | SAD: 60.7 -> 58.0 | Improved |
| add ASPP context (ablation) | UNet MobileNetv2 with indices and concat · Composition-1k | SAD: 54.7 -> 54.9 | No clear effect |
| replace max-pooling indices with nonlinear + context M2O depthwise IndexNet | B11 MobileNetv2 · Composition-1k | SAD: 54.9 -> 45.8 | Improved |
| replace holistic max index with nonlinear + context HIN | HMI baseline · Composition-1k | SAD: 56.5 -> 49.5 | Improved |
| add nonlinearity to linear holistic index network (ablation) | HIN · Composition-1k | SAD: 55.1 -> 50.6 | Improved |
| add context (4x4 kernel) to nonlinear holistic index network (ablation) | HIN · Composition-1k | SAD: 50.6 -> 49.5 | Improved |
| add nonlinearity to linear O2O depthwise index network with context (ablation) | O2O DIN · Composition-1k | SAD: 47.8 -> 50.2 | Hurt |
| increase backbone width multiplier to 1.4 instead of IndexNet (ablation) | B11 MobileNetv2 · Composition-1k | SAD: 54.9 -> 55.6 | No clear effect |
Closest papers by meaning
- Learning Affinity-Aware Upsampling for Deep Image Matting (CVPR 2021 · similarity 0.90)
- Deep Image Matting (CVPR 2017 · similarity 0.88)
- Revisiting Context Aggregation for Image Matting (ICML 2024 · similarity 0.86)
- Training Matting Models Without Alpha Labels (AAAI 2025 · similarity 0.86)
- Privileged Prior Information Distillation for Image Matting (AAAI 2024 · similarity 0.85)
- Path-Adaptive Matting for Efficient Inference Under Various Computational Cost Constraints (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.
