Source-linked AI summary
PseCo: Pseudo Labeling and Consistency Training for Semi-Supervised Object Detection
Gang Li, Xiang Li, Yujie Wang, Yichao Wu, Ding Liang, Shanshan Zhang
TL;DR
SSOD methods often overlook localization quality in pseudo boxes and feature-level consistency across scales. PseCo combines noisy pseudo-box learning with multi-view label- and feature-level consistency, outperforming prior methods on COCO while halving training time relative to Soft Teacher.
Problem
Existing SSOD pseudo labeling emphasizes classification scores without guaranteeing box localization, while random-resize consistency training neglects feature-level scale consistency.
Method
PseCo combines Noisy Pseudo box Learning, using prediction-guided label assignment and positive-proposal consistency voting, with Multi-view Scale-invariant Learning for label- and feature-level consistency.
Results
PseCo surpasses state-of-the-art methods on COCO and improves a 26.9% mAP baseline to 36.1% under a 10% labeling ratio.
Takeaways & Limitations
PseCo improves SSOD accuracy and efficiency, halving Soft Teacher’s training time while achieving better performance.
Abstract
from arXiv · showhide
In this paper, we delve into two key techniques in Semi-Supervised Object Detection (SSOD), namely pseudo labeling and consistency training. We observe that these two techniques currently neglect some important properties of object detection, hindering efficient learning on unlabeled data. Specifically, for pseudo labeling, existing works only focus on the classification score yet fail to guarantee the localization precision of pseudo boxes; For consistency training, the widely adopted random-resize training only considers the label-level consistency but misses the feature-level one, which also plays an important role in ensuring the scale invariance. To address the problems incurred by noisy pseudo boxes, we design Noisy Pseudo box Learning (NPL) that includes Prediction-guided Label Assignment (PLA) and Positive-proposal Consistency Voting (PCV). PLA relies on model predictions to assign labels and makes it robust to even coarse pseudo boxes; while PCV leverages the regression consistency of positive proposals to reflect the localization quality of pseudo boxes. Furthermore, in consistency training, we propose Multi-view Scale-invariant Learning (MSL) that includes mechanisms of both label- and feature-level consistency, where feature consistency is achieved by aligning shifted feature pyramids between two images with identical content but varied scales. On COCO benchmark, our method, termed PSEudo labeling and COnsistency training (PseCo), outperforms the SOTA (Soft Teacher) by 2.0, 1.8, 2.0 points under 1%, 5%, and 10% labelling ratios, respectively. It also significantly improves the learning efficiency for SSOD, e.g., PseCo halves the training time of the SOTA approach but achieves even better performance. Code is available at https://github.com/ligang-cs/PseCo.
1 Introduction
Semi-supervised object detection adapts pseudo labeling and consistency training but must account for localization quality and large variation in object scales. PseCo addresses these gaps with robust noisy-box learning and joint label- and feature-level scale consistency, improving COCO accuracy and training efficiency.
- 1 Introduction: SSOD methods inherit pseudo labeling and consistency training from classification while neglecting detection-specific localization and scale-invariance requirements.Detection pseudo labels contain both category and box information, and detectors must handle targets across a wide range of scales.
- 1 Introduction: Classification confidence does not guarantee pseudo-box localization: precision falls from 81% at IoU=0.3 to 31% at IoU=0.9.Noisy boxes can cause IoU-based assignment to label background proposals as foreground and confuse foreground-background decision boundaries.
- 1 Introduction: PseCo’s Noisy Pseudo box Learning uses prediction-guided assignment and positive-proposal consistency voting to make pseudo-box supervision robust to localization noise.Its Multi-view Scale-invariant Learning jointly learns label- and feature-level consistency by aligning shifted feature pyramids across differently scaled views.
- 1 Introduction: PseCo improves a 26.9% mAP baseline to 36.1% under a 10% labeling ratio, surpassing previous methods by at least 2.0%.With the full COCO training set labeled and 123K extra images unlabeled, it reaches 46.1% mAP from a 41.0% baseline, a +5.1% improvement.
- 1 Introduction: PseCo halves Soft Teacher’s training time while achieving better performance, improving SSOD learning efficiency.
2 Related Works
Prior semi-supervised learning methods center on pseudo labeling, consistency training, or their combination, with object-detection methods adapting these ideas to unlabeled images. Subsequent SSOD work improves pseudo-label quality and training frameworks, but the surveyed approaches remain extensions of these core paradigms.
- 2 Related Works: Semi-supervised image classification methods use pseudo labeling, consistency training, or combinations such as FixMatch and FlexMatch.
- 2 Related Works: STAC introduced pseudo labeling and consistency training for SSOD with strong augmentations but used two training stages that prevent pseudo-label updates during training.
- 2 Related Works: Later SSOD methods adopted EMA teacher updates for end-to-end pseudo-label generation, while others used model ensembles to improve pseudo-label quality and reduce confirmation bias.
3 Method
PseCo combines Noisy Pseudo box Learning with Multi-view Scale-invariant Learning to improve supervision from unlabeled images in semi-supervised object detection. NPL addresses noisy localization through prediction-guided assignment and consistency-based weighting, while MSL combines label- and feature-level consistency across scaled views.
- 3 Method: PseCo combines Noisy Pseudo box Learning (NPL) and Multi-view Scale-invariant Learning (MSL) for training on unlabeled data.The framework uses teacher-generated pseudo labels and trains the student on multiple views.
- 3.1 The basic framework: High pseudo-label thresholds improve category precision but reduce pseudo-label supply, worsening foreground-background and foreground-foreground imbalance.PseCo adopts focal loss in the basic framework to address these imbalance issues.
- 3.2 Noisy Pseudo Box Learning: PLA jointly uses teacher classification confidence and localization precision to rank proposals, reducing reliance on noisy pseudo-box IoUs.Proposal quality is defined as q = s^α × u^(1−α), and top proposals within candidate bags become positives.
- 3.2 Noisy Pseudo Box Learning: Positive-proposal Consistency Voting averages proposal-level IoUs and uses the resulting σ_j as an instance-wise regression-loss weight.The consistency score is positively correlated with true pseudo-box IoUs, supporting its use as a localization-quality estimate.
- 3.3 Multi-view Scale-invariant Learning: MSL trains on two student views: random resizing provides label consistency, while downsampling aligns shifted FPN levels for feature consistency.V1 and V2 use the same pseudo boxes; P3–P7 features of V1 align with P2–P6 features of V2, with 2x downsampling as default.
- 3.3 Multi-view Scale-invariant Learning: Unlike SoCo, MSL models both label and dense image-feature consistency, providing more comprehensive supervisory signals for scale-invariant learning.The method aligns dense features from shifted pyramids between paired images with identical content at different scales.
4 Experiments
On COCO, PseCo consistently outperforms strong semi-supervised object-detection baselines and converges faster, while ablations support both scale-invariant consistency and noisy-box learning components.
- Comparison with State-of-the-Art Methods: PseCo surpasses state-of-the-art methods by at least 1.7% across all partially labeled ratios, including 1.8% and 2.0% over Soft Teacher at 5% and 10%.At 1% and 2%, PseCo reaches 22.4 and 27.8 mAP, respectively.
- Comparison with State-of-the-Art Methods: PseCo reaches the same performance as Soft Teacher using only 2/5 and 1/4 of its iterations at 10% and 5% labeling ratios, respectively.The extra V2 view increases per-iteration training time by 25%, yet the overall method halves Soft Teacher’s training time.
- Comparison with State-of-the-Art Methods: 46.1 mAP represents a +5.1% improvement over the 41.0 mAP supervised baseline under the Fully Labeled Data setting.PseCo establishes a new state of the art in this setting.
- Ablation Study: PCV outperforms regression-consistency and box-jittering alternatives by 0.9% and 0.5% on AP75, respectively.PCV uses intrinsic proposals, avoids extra forward passes on jittered boxes, and requires no variance hyperparameter.
- Ablation Study: Adding feature-level consistency to label-level consistency improves performance by +1.0%, reaching 34.9 mAP, and also accelerates convergence.The label-consistency view alone surpasses single-scale training by 1.2 mAP; the feature-consistency view provides the further gain.
- Ablation Study: Vanilla multi-view training with only label consistency remains at 33.9 mAP, matching single-view training and underscoring the role of feature consistency.The comparison uses an additional downsampled view that performs label consistency like V1.
5 Conclusion
PseCo integrates object-detection-specific handling into pseudo labeling and consistency training through NPL and MSL. Experiments on COCO show that it surpasses state-of-the-art methods in accuracy and efficiency.
- NPL uses prediction-guided label assignment and positive-proposal consistency voting to robustly handle noisy pseudo boxes.
- MSL combines label-level consistency with feature-level scale-invariant learning, addressing a property neglected by prior work.
- PseCo surpasses state-of-the-art methods by a large margin in both accuracy and efficiency on the COCO benchmark.