Research map / Label-efficient and robust learning / Semi-supervised learning
Detect Any Mirrors: Boosting Learning Reliability on Large-Scale Unlabeled Data with an Iterative Data Engine
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.
Collects a 0.4M-image unlabeled mirror dataset and proposes DAM, an iterative data engine that filters pseudo labels with geometric and multi-modal-agent semantic scores before reliability-enhanced mean-teacher training for mirror detection.
Problem
Semi-supervised mirror detection with large unlabeled data suffers from unreliable teacher pseudo-labels caused by distribution shift between labeled and unlabeled data, and limited annotated mirror data hurts generalization.
Approach
An iterative data engine that scores each pseudo label with a learned geometric-accuracy estimator (predicting IoU, Fβ, Accuracy) and a learned semantic scorer distilled from MLLM agents (MiniCPM-V-2.6, Qwen2-VL, InternVL-2 aggregated by GLM4-flash), keeps only high-scoring samples, and trains a mean-teacher model on them.
Key results
- DAM reaches IoU 0.934 on MSD, 0.734 on PMD and 0.808 on RGBD-Mirror*, beating all 11 compared methods on all five metrics.
- Without training on VMD-D or MMD, DAM achieves IoU 0.755 on VMD-D (vs. 0.723 for DepthAnything) and 0.759 on MMD (vs. 0.732).
- Ablation on PMD: supervised baseline IoU 0.679 rises to 0.734 with geometric scoring, semantic scoring and reliability-enhanced training combined.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add geometric scoring to the labeled-data baseline (ablation) | DAM · PMD | IoU: 0.679 -> 0.693 | Helped |
| add MMA-driven semantic scoring to geometric scoring (ablation) | DAM · PMD | IoU: 0.693 -> 0.719 | Helped |
| add reliability-enhanced training to geometric and semantic scoring (ablation) | DAM · PMD | IoU: 0.719 -> 0.734 | Helped |
Surprising findings
- Semi-supervised baselines such as UniMatch underperform the supervised SegFormer/Mask2Former on PMD in IoU (0.698 vs 0.698 and 0.706 for MT), suggesting unreliable pseudo labels can hurt.
Limitations
- MLLM-based semantic scoring is time-consuming and must be distilled into a lightweight scorer.
- Pseudo-label quality depends on the initial model trained on limited annotated data.
- Selection threshold T must be tuned empirically (T=0.6 chosen).
Experimental setup
Datasets: MSD PMD RGBD-Mirror VMD-D MMD
Benchmarks: MSD PMD RGBD-Mirror* VMD-D MMD
Metrics: IoU Accuracy F-beta MAE BER
Baselines: SegFormer Mask2Former MirrorNet PMDNet SANet VCNet HetNet CSFwinformer MT UAMT DepthAnything UniMatch
Builds on: Mean Teacher UAMT UniMatch DepthAnything DINOv2 DPT
Closest papers by meaning
- Weakly-Supervised Mirror Detection via Scribble Annotations (AAAI 2024 · similarity 0.91)
- ZOOM: Learning Video Mirror Detection with Extremely-Weak Supervision (AAAI 2024 · similarity 0.91)
- Effective Video Mirror Detection with Inconsistent Motion Cues (CVPR 2024 · similarity 0.86)
- Dual-Perspective Knowledge Enrichment for Semi-supervised 3D Object Detection (AAAI 2024 · similarity 0.84)
- Automated Detection of Visual Attribute Reliance with a Self-Reflective Agent (NeurIPS 2025 · similarity 0.84)
- Segment Anything without Supervision (NeurIPS 2024 · similarity 0.83)
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.