Research map / Detection and segmentation / Weakly supervised detection and segmentation
Weakly-Supervised Mirror Detection via Scribble Annotations
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 the first weakly-supervised mirror detection framework trained on a new scribble-annotated mirror dataset (S-Mirror, 10,158 images), using LGFE, FAMA and prototype contrast loss on a PVT backbone.
Problem
Mirror detection normally relies on dense pixel-level labels, which are expensive; weak supervision for mirror regions (with scale variation, occlusion, and reflection interference) is unexplored.
Approach
Scribble-based annotation of mirror and non-mirror regions (under 0.01 labeled pixels, about 8 s per image) combined with a local-global feature enhancement module, a foreground-aware mask attention fusing semantic and edge cues, and a cross-image prototype contrast loss.
Key results
- On MSD, the weakly supervised method reaches S-measure 0.828 and MAE 0.078, outperforming SS, SCWS and WSCOD and approaching fully supervised MD models.
- On Mirror-RGBD, S-measure 0.754 versus 0.698 for the best weakly supervised baseline WSCOD.
- Ablation on MSD shows LGFE, FAMA and PCL each improve over baseline, with the full model reaching S-measure 0.828 from 0.793.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the proposed weakly supervised method instead of WSCOD | Mirror detection · MSD | S-measure (Sm): 0.786 -> 0.828 | Helped |
| Use the proposed weakly supervised method instead of WSCOD | Mirror detection · PMD | S-measure (Sm): 0.764 -> 0.773 | Helped |
| Use the proposed weakly supervised method instead of WSCOD | Mirror detection · Mirror-RGBD | S-measure (Sm): 0.698 -> 0.754 | Helped |
| Add LGFE (ablation) | Baseline mirror detector · MSD | S-measure (Sm): 0.793 -> 0.810 | Helped |
| Add FAMA (ablation) | Baseline mirror detector · MSD | S-measure (Sm): 0.793 -> 0.818 | Helped |
| Add FAMA after LGFE (ablation) | Baseline + LGFE · MSD | S-measure (Sm): 0.810 -> 0.820 | Helped |
| Add PCL after LGFE and FAMA (ablation) | Baseline + LGFE + FAMA · MSD | S-measure (Sm): 0.820 -> 0.828 | Helped |
Surprising findings
- The weakly supervised method outperforms several fully supervised SOD models (e.g., CPDNet, MINet, R3Net) on MSD, though it trails the best fully supervised MD models.
Limitations
- Relies on scribble annotations and edge maps from Canny, so supervision is still noticeable compared with image-level labels.
- Still falls short of the best fully supervised mirror detection models such as SATNet and HetNet on MSD and PMD.
Experimental setup
Datasets: MSD PMD Mirror-RGBD S-Mirror
Benchmarks: MSD PMD Mirror-RGBD
Metrics: S-measure E-measure weighted F-measure IoU MAE
Baselines: CPDNet MINet LDF VST R3Net EGNet PoolNet SETR MirrorNet PMDNet HetNet SATNet
Builds on: PVT CFM DenseASPP MDTA scribble-based SOD (Zhang et al. 2020)
Closest papers by meaning
- ZOOM: Learning Video Mirror Detection with Extremely-Weak Supervision (AAAI 2024 · similarity 0.93)
- Detect Any Mirrors: Boosting Learning Reliability on Large-Scale Unlabeled Data with an Iterative Data Engine (CVPR 2025 · similarity 0.91)
- Progressive Mirror Detection (CVPR 2020 · similarity 0.89)
- Learning To Detect Mirrors From Videos via Dual Correspondences (CVPR 2023 · similarity 0.88)
- Scribble Hides Class: Promoting Scribble-Based Weakly-Supervised Semantic Segmentation with Its Class Label (AAAI 2024 · similarity 0.85)
- Effective Video Mirror Detection with Inconsistent Motion Cues (CVPR 2024 · 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.