Research map / Image generation / Autoregressive and latent generation
Privileged Prior Information Distillation for Image Matting
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes PPID-IM, which distills privileged trimap-based teacher knowledge (via cross-level semantic and attention-guided local distillation) into trimap-free image matting students, and introduces the Real-1K UHD matting benchmark.
Problem
Trimap-free image matting struggles to decouple deterministic and undetermined regions when foregrounds are semantically ambiguous, chromaless, or highly transmissive.
Approach
Treat the trimap as privileged information available only to a trimap-based teacher with the same architecture as the trimap-free student; transfer its semantic features across layers (CLSD, using GC blocks) and its attention-weighted local transition-region features (ALD) to the student, with the trimap unused at inference.
Key results
- On Real-1K whole test set, IndexNet-PPID gives the largest gains among baselines, with MSE reduced by about 61.8% on transparent and 64.0% on non-transparent subsets relative to the trimap-free IndexNet baseline (as stated in the text).
- PPID-guided trimap-free models (e.g., LFM-PPID, Hatt-PPID, AIM-PPID) improve over their trimap-free counterparts on the Real-1K transparent and non-transparent groups.
- Ablations on Composition-1K show CLSD outperforms same-layer semantic distillation (SD), and ALD with attention outperforms plain local distillation (LD) on transition regions.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add PPID to trimap-free models | trimap-free matting baselines · Real-1K | MSE: 0.047 -> 0.026 | Helped |
| add PPID to trimap-free models | trimap-free matting baselines · Real-1K transparent images | MSE: 0.068 -> 0.026 | Helped |
| add PPID to trimap-free models | trimap-free matting baselines · Real-1K non-transparent images | MSE: 0.025 -> 0.009 | Helped |
| use cross-layer semantic distillation instead of same-layer semantic distillation (ablation) | DIM · Adobe Composition-1K | MSE: 0.016 -> 0.014 | Helped |
| use cross-layer semantic distillation instead of same-layer semantic distillation (ablation) | IndexNet · Adobe Composition-1K | SAD: 59.37 -> 48.35 | Helped |
| use cross-layer semantic distillation instead of same-layer semantic distillation (ablation) | GCA · Adobe Composition-1K | SAD: 62.64 -> 57.27 | Helped |
| add same-layer semantic distillation (ablation) | DIM · Adobe Composition-1K | SAD: 96.92 -> 92.61 | Helped |
| add same-layer semantic distillation (ablation) | IndexNet · Adobe Composition-1K | SAD: 67.12 -> 59.37 | Helped |
| add same-layer semantic distillation (ablation) | GCA · Adobe Composition-1K | SAD: 72.02 -> 62.64 | Helped |
| add local distillation without attention guidance (ablation) | DIM · AIM transition region | SAD: 120.08 -> 60.19 | Helped |
| add attention-guided local distillation instead of local distillation without attention guidance (ablation) | DIM · AIM transition region | SAD: 60.19 -> 51.68 | Helped |
| add local distillation without attention guidance (ablation) | IndexNet · AIM transition region | SAD: 70.59 -> 58.72 | Helped |
Surprising findings
- Trimap-free students distilled with PPID sometimes surpass their trimap-based teachers on real-world images, while the original trimap-free methods do not.
Limitations
- Trained on a limited amount of composited data, so real-world generalization relies on distillation effects.
- Teacher and student share identical architecture, so benefit for architectures with different structures is untested.
Experimental setup
Datasets: Adobe Composition-1K Distinction-646
Benchmarks: Real-1K Adobe Composition-1K Distinction-646
Metrics: SAD MSE Grad Conn
Baselines: DIM IndexNet GCA Matting MatteFormer LFM Hattt (Distinction-646 Hatt) AIM
Builds on: generalized distillation (learning using privileged information) GCNet global context block CBAM Deep Image Matting
Closest papers by meaning
- Improved Image Matting via Real-Time User Clicks and Uncertainty Estimation (CVPR 2021 · similarity 0.89)
- Training Matting Models Without Alpha Labels (AAAI 2025 · similarity 0.88)
- Unifying Automatic and Interactive Matting with Pretrained ViTs (CVPR 2024 · similarity 0.87)
- DRIP: Unleashing Diffusion Priors for Joint Foreground and Alpha Prediction in Image Matting (NeurIPS 2024 · similarity 0.87)
- In-Context Matting (CVPR 2024 · similarity 0.86)
- Revisiting Context Aggregation for Image Matting (ICML 2024 · similarity 0.86)
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.