Research map / Detection and segmentation / Weakly supervised detection and segmentation
Multi-clue Consistency Learning to Bridge Gaps Between General and Oriented Object in Semi-supervised Detection
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 Multi-clue Consistency Learning (MCL), a pixel-level pseudo-labeling semi-supervised framework for oriented object detection on aerial images that adds Gaussian Center Assignment, Scale-aware Label Assignment, and Consistent Confidence Soft Label to a FCOS-based Dense Teacher pipeline.
Problem
Semi-supervised object detection methods designed for general scenes degrade on oriented objects in aerial images due to sampling, assignment, and confidence inconsistencies (large aspect ratios, noisy rotated pseudo-boxes, and mismatched classification/localization confidence).
Approach
Replace center sampling with Gaussian-shaped positive assignment for oriented boxes, select pixel-level pseudo-labels per FPN level group (coarse-to-fine for small scales, score threshold for large scales) instead of pseudo-boxes, and supervise classification with a centerness-based soft label scaled by object size.
Key results
- On DOTA-v1.5-Partial, MCL reaches 52.98, 59.63 and 62.63 mAP at 10%, 20%, 30% labeled, exceeding the prior SOOD state of the art PST by roughly 2 to 3.5 mAP.
- On DOTA-v1.5-Full, MCL reaches 69.08 mAP (+3.62 over the supervised FCOS baseline of 65.46), versus 67.70 for SOOD*.
- On DOTA-v1.0-Full, MCL reaches 73.63 mAP (+6.91 over supervised 66.72), versus 70.81 for SOOD*.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply MCL instead of PST | FCOS-based semi-supervised oriented detection · DOTA-v1.5-Partial 10% | mAP(%): 49.63 -> 52.98 | Helped |
| Apply MCL instead of PST | FCOS-based semi-supervised oriented detection · DOTA-v1.5-Partial 20% | mAP(%): 57.39 -> 59.63 | Helped |
| Apply MCL instead of PST | FCOS-based semi-supervised oriented detection · DOTA-v1.5-Partial 30% | mAP(%): 60.40 -> 62.63 | Helped |
| Apply MCL instead of SOOD* | FCOS-based semi-supervised oriented detection · DOTA-v1.5-Full | mAP: 67.70 -> 69.08 | Helped |
| Apply MCL instead of SOOD* | FCOS-based semi-supervised oriented detection · DOTA-v1.0-Full | mAP: 70.81 -> 73.63 | Helped |
| Add GCA (ablation) | Dense Teacher baseline · DOTA-v1.5-Partial 30% | mAP(%): 59.65 -> 60.62 | Helped |
| Add SLA (ablation) | Dense Teacher baseline · DOTA-v1.5-Partial 30% | mAP(%): 59.65 -> 61.32 | Helped |
| Add CCSL (ablation) | MCL with GCA and SLA · DOTA-v1.5-Partial 30% | mAP(%): 61.46 -> 62.63 | Helped |
| Replace center sampling with all pixels inside ground-truth boxes (ablation) | FCOS · DOTA-v1.5-Partial 30% | mAP: 59.65 -> 58.49 | Hurt |
| Use pixel-level pseudo-labels instead of pseudo-boxes (ablation) | FCOS semi-supervised detector · DOTA-v1.5-Partial 30% | mAP: 26.02 -> 59.65 | Helped |
| Use joint score × centerness selection instead of score selection at 3% (ablation) | Dense Teacher · DOTA-v1.5-Partial 30% | mAP: 59.65 -> 60.45 | Helped |
| Replace centerness quality target with IoU (ablation) | CCSL · DOTA-v1.5-Partial 30% | mAP: 60.98 vs 62.63 | Hurt |
Surprising findings
- Plain Mean Teacher with pseudo-boxes drops to 26.02 mAP on DOTA-v1.5 (from 54.79 supervised FCOS), while pixel-level pseudo-labels (Dense Teacher) reach 59.65.
- All-pixel sampling inside ground-truth boxes performs worse than center sampling (58.49 vs 59.65 mAP), attributed to excessive background at edges.
Limitations
- Sensitivity to hyperparameters such as ratio, top-k, threshold and beta is evident in ablations; settings were tuned on DOTA.
- Evaluation limited to DOTA benchmarks and a single FCOS-based detector backbone (ResNet-50).
Experimental setup
Datasets: DOTA-v1.5 DOTA-v1.0 COCO
Benchmarks: DOTA-v1.5 DOTA-v1.0
Metrics: mAP pixel-level recall pixel-level precision
Baselines: Faster RCNN FCOS Unbiased Teacher Soft Teacher Dense Teacher PseCo DualPolish ARSL SOOD PST
Builds on: Dense Teacher FCOS Mean Teacher SOOD
Closest papers by meaning
- Consistency-based Semi-supervised Learning for Object detection (NeurIPS 2019 · similarity 0.91)
- Semi-DETR: Semi-Supervised Object Detection With Detection Transformers (CVPR 2023 · similarity 0.90)
- Ambiguity-Resistant Semi-Supervised Learning for Dense Object Detection (CVPR 2023 · similarity 0.90)
- Rethinking Pseudo Labels for Semi-supervised Object Detection (AAAI 2022 · similarity 0.90)
- Relational Matching for Weakly Semi-Supervised Oriented Object Detection (CVPR 2024 · similarity 0.88)
- Sparse Semi-DETR: Sparse Learnable Queries for Semi-Supervised Object Detection (CVPR 2024 · similarity 0.87)
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.