Research map / Video / Video segmentation and tracking
Effective Video Mirror Detection with Inconsistent Motion Cues
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 MG-VMD, a video mirror detection network that models inconsistent inside/outside motion via optical flow (MAM and MEDM modules), and introduces the MMD video mirror dataset.
Problem
Video mirror detection is unreliable under prior dual-correspondence assumptions, and the existing VMD-D dataset is biased toward small, centrally placed mirrors, limiting real-world applicability.
Approach
Exploit the inconsistency between motion inside and outside a mirror (e.g., smaller motion inside due to depth perception) using optical flow, via an attention module that localizes mirrors from flow fields and a motion-guided cross-attention edge module, fused with a PMD-Net-style refinement.
Key results
- MG-VMD reaches Accuracy 0.872532, Fβ 0.869419, IoU 0.725130 on MMD, the best among compared methods (VMD-Net: 0.854415 accuracy, 0.812347 Fβ).
- Ablation: adding MAM to the baseline raises accuracy from 0.857519 to 0.866218; adding MEDM gives 0.858176; full model reaches 0.872532.
- Reports that MMD has higher SSIM between train and test splits (0.903) than VMD-D (0.795), indicating a fairer split.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use MG-VMD instead of VMDNet | Video mirror detection · MMD test | Accuracy: 0.854415 -> 0.872532 | Helped |
| Use MG-VMD instead of VMDNet | Video mirror detection · MMD test | MAE: 0.145585 -> 0.127468 | Helped |
| Use MG-VMD instead of FSNet | Video mirror detection · MMD test | Fβ: 0.837732 -> 0.869419 | Helped |
| Use MG-VMD instead of VMDNet | Video mirror detection · MMD test | IoU: 0.722634 -> 0.725130 | Helped |
| Add MEDM to the baseline (ablation) | MG-VMD · MMD test | Accuracy: 0.857519 -> 0.858176 | Helped |
| Add MEDM to the baseline (ablation) | MG-VMD · MMD test | Fβ: 0.837578 -> 0.847853 | Helped |
| Add MEDM to the baseline (ablation) | MG-VMD · MMD test | IoU: 0.700855 -> 0.707350 | Helped |
| Add MAM to the baseline (ablation) | MG-VMD · MMD test | Accuracy: 0.857519 -> 0.866218 | Helped |
| Add MAM to the baseline (ablation) | MG-VMD · MMD test | MAE: 0.142481 -> 0.133782 | Helped |
| Add MAM to the baseline (ablation) | MG-VMD · MMD test | IoU: 0.700855 -> 0.713005 | Helped |
| Add both MEDM and MAM to the baseline (ablation) | MG-VMD · MMD test | Fβ: 0.837578 -> 0.869419 | Helped |
| Add both MEDM and MAM to the baseline (ablation) | MG-VMD · MMD test | IoU: 0.700855 -> 0.725130 | Helped |
Surprising findings
- Window regions with irregular motion are often over-predicted as mirrors, a failure mode of motion cues.
Limitations
- Over-predicts windows as mirrors because windows also show irregular motion.
- Dataset is small: 37 videos from six participants, with only 18 training videos.
Experimental setup
Datasets: MMD VMD-D
Benchmarks: MMD
Metrics: MAE F-measure Accuracy IoU
Baselines: VMD-Net PMD-Net MirrorNet UFO FSNet F3Net MGA ALGRF
Builds on: VMD-Net PMD-Net DeepLabV3 ResNeXt-101
Closest papers by meaning
- ZOOM: Learning Video Mirror Detection with Extremely-Weak Supervision (AAAI 2024 · similarity 0.89)
- Detect Any Mirrors: Boosting Learning Reliability on Large-Scale Unlabeled Data with an Iterative Data Engine (CVPR 2025 · similarity 0.86)
- Weakly-Supervised Mirror Detection via Scribble Annotations (AAAI 2024 · similarity 0.85)
- MIRROR: Make Your Object-Level Multi-View Generation More Consistent with Training-Free Rectification (ICML 2025 · similarity 0.84)
- Multi-View Dynamic Reflection Prior for Video Glass Surface Detection (AAAI 2024 · similarity 0.84)
- M^3-VOS: Multi-Phase, Multi-Transition, and Multi-Scenery Video Object Segmentation (CVPR 2025 · similarity 0.82)
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.