Research map / Architectures and efficiency / CNN design
Learning Affinity-Aware Upsampling for Deep Image Matting
Yutong Dai, Hao Lu, Chunhua Shen
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 |
|---|---|---|---|
| replace nearest-neighbor upsampling with affinity-aware dynamic-cs-d upsampling | ResNet34 matting network · Composition-1k test | SAD: 37.51 -> 32.15 | Improved |
| replace nearest-neighbor upsampling with paired affinity-aware downsampling and upsampling | ResNet34 matting network · Distinctions-646 test | SAD: 25.03 -> 23.20 | Improved |
| use affinity-aware upsampling instead of IndexNet | MaxPool reconstruction network · MNIST | PSNR: 45.51 -> 47.63 | Improved |
| use affinity-aware upsampling instead of IndexNet | MaxPool reconstruction network · Fashion-MNIST | PSNR: 45.83 -> 46.41 | Improved |
| pair dynamic-cs downsampling with upsampling (ablation) | A²U matting network · Composition-1k test | SAD: 35.86 -> 33.13 | Improved |
| add normalization and nonlinearity after the encoding layers (ablation) | A²U dynamic-cs-d matting network · Composition-1k test | SAD: 33.13 -> 32.15 | Improved |
| use channel-wise rather than channel-shared encoding in the hybrid implementation (ablation) | A²U matting network · Composition-1k test | SAD: 36.43 -> 34.76 | Improved |
| use channel-shared rather than channel-wise encoding in the dynamic implementation (ablation) | A²U matting network · Composition-1k test | SAD: 36.66 -> 35.86 | Improved |
| set the upsampling kernel size to 3 rather than 1 (ablation) | A²U hybrid-cw matting network · Composition-1k test | SAD: 37.74 -> 34.76 | Improved |
| increase the upsampling kernel size from 3 to 5 (ablation) | A²U hybrid-cw matting network · Composition-1k test | SAD: 34.76 -> 35.99 | Hurt |
| use sigmoid+softmax instead of softmax normalization (ablation) | A²U hybrid-cw matting network · Composition-1k test | SAD: 35.93 -> 34.76 | Improved |
| use sigmoid+softmax instead of softmax normalization (ablation) | A²U dynamic-cs matting network · Composition-1k test | SAD: 36.40 -> 35.86 | Improved |
Closest papers by meaning
- Indices Matter: Learning to Index for Deep Image Matting (ICCV 2019 · similarity 0.90)
- FADE: Fusing the Assets of Decoder and Encoder for Task-Agnostic Upsampling (ECCV 2022 · similarity 0.88)
- Boosting Robustness of Image Matting With Context Assembling and Strong Data Augmentation (CVPR 2022 · similarity 0.88)
- Revisiting Context Aggregation for Image Matting (ICML 2024 · similarity 0.87)
- Training Matting Models Without Alpha Labels (AAAI 2025 · similarity 0.85)
- Path-Adaptive Matting for Efficient Inference Under Various Computational Cost Constraints (AAAI 2025 · 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.
