Aligned Consensus Teaching for Label-Efficient Oriented Object Detection in Weakly-Aligned Visible-Infrared Imagery
Qi Ming, Xiaxin Yuan, Jiahuan Zhou, Jiangmeng Li, Xudong Zhao, Zhanchao Huang, Juan Fang, Shaoguang Huang, et al.
cs.CV
Quick verdict
First semi-supervised framework for image-pair-level RGB–IR detection; strong results, but unmatched baselines and single-run peaks temper headline gains.
Summary
The paper introduces image-pair-level semi-supervised visible-infrared object detection (VIOD), where only a small fraction of RGB-IR pairs carry boxes in both modalities and the rest are fully unlabeled, and proposes ACT, presented as the first framework for this setting. ACT combines three mechanisms: Cycle-Consistent Region Alignment (CRA) regularizes cross-modal correspondence on labeled pairs via target-free cycle consistency, sparse anchors, and reliability-weighted matching; Cross-Modal Consensus Mean-Teacher (CMC-MT) generates pseudo labels from the fused branch under geometry-preserving weak/strong views to recover branch-wise misses; Text-Guided Cross-Modal Instance Augmentation (TG-CMIA) composes tail-class RGB-IR instance pairs guided by a frozen Qwen2.5-VL-32B scene prior while preserving cross-modal offsets. On DroneVehicle, ACT reaches 0.511/0.644/0.683 mAP50 at 1/5/10% budgets, exceeding the strongest semi-supervised baseline by up to 6.2 points and recovering 94.3% of full-supervision mAP (0.724) at 10%. On VEDAI it reaches 0.541-0.632 with 10+ point margins. Limitations: CRA contributes only 0.1-0.5 mAP in the sequential ablation, headline tables are single-run peaks over baselines running on unmatched architectures, VEDAI shows near-flat budgets without variance estimates, and evaluation uses validation rather than the official DroneVehicle test split.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
5 strengths · 5 weaknesses · 5 author questions
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- ACT is the first framework for image-pair-level semi-supervised visible-infrared oriented object detection, where few pairs are labeled in both modalities and the rest are fully unlabeled.
- Cross-modal consensus pseudo labels from the fused branch recover branch-wise misses, improving pseudo-label recall and precision over either single-modality branch (Fig. 3b).
- On DroneVehicle, 10% labeled pairs suffice to reach 94.3% of full-supervision mAP50 (0.683 vs 0.724), and 1% labeled pairs still yield 0.511 vs 0.299 for the supervised baseline.
- TG-CMIA composes tail-class RGB-IR instance pairs that preserve cross-modal offsets, raising the target-class mean AP from 0.252 to 0.276 at the 1% budget.
- ACT consistently outperforms both supervised and semi-supervised state-of-the-art methods on DroneVehicle and VEDAI across all tested annotation budgets.
Strengths
Well-motivated new task setting with quantified evidence
The image-pair-level protocol is concretely grounded: at a 1% budget DroneVehicle offers 3.2k vs 286.8k matched boxes (~90x fewer) for alignment supervision, and branch recall asymmetry (RGB 73.2% vs IR 91.1% at 10%) directly motivates consensus pseudo labels (Fig. 1).
Unusually transparent ablation reporting
Sec 4.4 explicitly states the sequential ablation yields conditional rather than isolated causal gains, and reports CRA's small contribution (0.1/0.5/0.4 points) rather than hiding it; per-component effects are decomposed by budget and class (Fig. 3c: target-class mean 0.252->0.276 at 1%).
Beyond-standard robustness checks
Diagnostic evidence for the core mechanism
Inference-cost neutrality is explicitly engineered and reported
Weaknesses
Baseline architectural fairness is unresolved
Headline results are single-run peaks
CRA, a headline contribution, shows marginal and oddly ordered gains
VEDAI protocol and stability concerns
Ungoverned VLM dependency in TG-CMIA
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What does image-pair-level semi-supervised visible-infrared detection mean in this paper?
It is a training setting where only a small percentage of paired visible (RGB) and infrared images — for example 1-10% — have oriented bounding-box annotations in both modalities, while all remaining pairs are completely unlabeled. The detector must learn from these few jointly labeled pairs plus many unlabeled pairs, instead of the fully supervised setting where every pair is annotated twice. The authors state ACT is the first framework designed for this specific setting.
How much labeled data does ACT need to approach full supervision?
On DroneVehicle, with 10% of image pairs labeled (1,799 pairs), ACT reaches 0.683 mAP at IoU 0.5, which is 94.3% of the 0.724 achieved by the same detector trained on 100% of labels. Even at the extreme 1% budget (180 labeled pairs) it reaches 0.511, versus 0.299 for the same architecture trained supervised-only on that subset.
Does ACT add inference-time cost compared to a normal detector?
No. All three proposed components (alignment regularization, consensus pseudo-labeling, and text-guided instance augmentation) operate only during training. At inference the model retains only the EMA teacher's fused branch and the oriented detection head: 120.83M parameters running at 3.2 FPS (batch size 1, 512x640 on an A800 GPU), matching the supervised baseline's deployment footprint.
Which datasets and metrics are used to evaluate ACT?
Experiments use DroneVehicle (28,439 RGB-IR pairs, 5 vehicle classes, official 17,990/1,469/8,980 train/validation/test split) and VEDAI (772 training and 192 validation pairs in the authors' split, 4 classes, average object width around 14 pixels). Evaluation reports rotated mAP at IoU 0.5 with 11-point interpolation against infrared annotations, with identical splits and evaluation code for all compared methods.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 1-4 are available only through extracted text descriptions; their visual content could not be inspected directly.
- The supplementary material (complete per-seed results, full run details, CRA weight analysis) is not included in the extracted text.
- This is an arXiv v1 preprint (16 Sep 2026) with no evidence of peer review; claims of firstness could not be independently verified within the text.
- Code availability is a promise ('will be available on GitHub') with no repository link in the text, so it could not be verified.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.