Rethinking Pseudo Labels for Semi-Supervised Object Detection

1 Introduction

• Collecting human annotations is labor-intensive and time-consuming, particularly for object detection, where annotations are more expensive to obtain.The introduction contrasts relatively limited work on semi-supervised object detection with extensive study of semi-supervised image classification.Source: §1 · p. 1
• Conventional classification-score thresholds do not reflect box localization quality and can filter lower-confidence predictions from underrepresented classes.The passage illustrates the class imbalance with 9 toasters and 12,343 persons in 5% of COCO training data.Source: §1 · p. 1
• The approach builds on teacher-student pseudo labeling and makes each pseudo label reflect both localization quality and classification confidence.The teacher generates boxes and class predictions on unlabeled images for use as student supervision.Source: §1 · p. 1
• Localization is formulated as classification over intervals, with maximal scores across four box sides averaged as a localization-quality measure, then refined within intervals.The combined localization and classification scores condition per-category thresholds and loss reweighting to mitigate class imbalance.Source: §1 · p. 2
• The method improves state-of-the-art performance by 1-2% on COCO and PASCAL VOC, and improves the supervised baseline by up to 10% AP with 1/2/5/10% COCO annotations.The reported improvements use various training recipes.Source: §1 · p. 2

2 Approach

• The method combines certainty-aware pseudo labels with class-conditioned thresholds and loss reweighting for semi-supervised object detection.It uses localization and classification certainty to guide pseudo-label generation.Source: §2 · p. 2

Preliminary

• A teacher trained on labeled images generates pseudo boxes for unlabeled images, which supervise a student trained with strong augmentations.Pseudo labels are generated using horizontal flipping; student training uses color jitter, Gaussian blur, and Cutout.Source: Preliminary · p. 3

Certainty-aware Pseudo Labels

• Candidate-box sides are localized through interval classification, producing certainty-aware boxes that account for both localization quality and classification confidence.Each side is assigned to one of K intervals using a K-way classification task; the maximum score per side is averaged to measure localization quality.Source: Certainty-aware Pseudo Labels · p. 3
• Fine regression from each interval center to the ground-truth side refines localization beyond the discretized interval estimate.The method uses Smooth L1 for this refinement and replaces the localization loss in both teacher and student models.Source: Certainty-aware Pseudo Labels · p. 3

Dynamic Thresholding and Re-weighting

• Class-specific thresholds filter pseudo labels, while class-specific loss weights promote underrepresented categories; γ1 and γ2 control the focus on those classes.The paper reports clipping thresholds to [0.4, 0.9] works well empirically.Source: Dynamic Thresholding and Re-weighting · p. 4

Main Results

• The method outperforms state-of-the-art by 1–2% AP on COCO and PASCAL VOC; with 10% COCO labels, it reaches 32.23% AP, about 2% above Instant-Teaching.The reported comparison includes full-data and limited-annotation settings.Source: Main Results · p. 5, Table 2 · p. 5
• The method reports state-of-the-art APs of 19.02%, 23.34%, and 28.40%, respectively; the cited passage does not identify which annotation settings correspond to these values.Source: Main Results · p. 5
• Adding dynamic thresholding and loss reweighting raises overall AP from 21.6% to 22.5% and rare-class AP from 23.9% to 26.0%.The method also gains 0.8% AP50 and 0.5% overall AP with model ensembling; 1:1 labeled-to-unlabeled sampling further improves performance.Source: Main Results · p. 5, Table 3 · p. 5, Table 4 · p. 5
• At 0.85 IoU, the method achieves more than 6% higher mAP than the baseline, while performance is similar at 0.5 IoU.The teacher-model pseudo-box evaluation follows a similar localization trend.Source: Figure 2 · p. 6, Main Results · p. 6, Figure 3 · p. 6
• Qualitative examples show more precise localization than the baseline, particularly around irregularly posed objects such as the bear and person.Source: Main Results · p. 6, Figure 4 · p. 7

Ablation Study

• Certainty-aware pseudo labels, loss reweighting, and dynamic thresholding improve AP by 1.7%, 0.8%, and 0.5%; together they improve AP by 2.6% and AP75 by 4.3%.Source: Table 6 · p. 6, Ablation Study · p. 6
• Localization is robust when K provides fine-grained intervals, but K=4 yields 19.7% AP; γ1=0.05 and γ2=0.6 give the best reported result for thresholding and reweighting.Larger γ1 and γ2 emphasize infrequent classes, while setting either factor to zero disables its corresponding mechanism.Source: Ablation Study · pp. 6–7
• Without student-model augmentation, AP falls from 22.5% to 20.3%; color jitter and Gaussian blur improve it by 1.2%, and Cutout adds 1%.Source: Table 7 · p. 6, Ablation Study · p. 6

Object Detection

• Object detectors have progressed from anchor-based designs to anchor-free and transformer-based approaches, while requiring box annotations for objects of interest.Source: §4 · p. 7

Semi-Supervised Learning

• Visual semi-supervised learning commonly uses consistency across augmented views, and Mean Teacher-style methods update a teacher with an exponential moving average of the student.Source: §4 · p. 7

Semi-Supervised Object Detection

• Existing SSOD methods adapt consistency-based approaches such as CSD and STAC, while this work targets localization precision and dynamic pseudo-label thresholds for class imbalance.Source: §4 · p. 7

• The method targets two SSOD issues: missing localization confidence in pseudo-label generation and amplified class imbalance.Source: §5 · p. 7
• It estimates both classification and localization quality by formulating box localization as a classification problem.Quality scores condition dynamically derived pseudo-label thresholds and class-specific loss reweighting.Source: §5 · p. 8
• Experiments under multiple settings demonstrate the efficacy of the certainty-aware pseudo-labeling method.Source: §5 · p. 8

Appendix Training and Testing Configuration

• COCO-full uses 180k/540k teacher and student iterations for 1×/3× schedules, while COCO-partial uses 6k/12k/30k/60k teacher iterations and 180k student iterations for 1/2/5/10% settings.Partial-setting evaluation uses score thresholds 0.05 and 0.001; the lower threshold generally improves recall and slightly improves performance.Source: Training and Testing Configuration · p. 8
• PASCAL VOC uses 10k teacher and 90k student iterations, with both single-scale and multi-scale results reported.Source: Training and Testing Configuration · p. 8
• COCO ablations use the 2% setting and a shorter 0.5× schedule, and the paper reports a total batch size of 8 versus 16/32/64 in existing methods.The authors characterize their schedule as equal to, and sometimes shorter than, other state-of-the-art SSOD approaches.Source: Training and Testing Configuration · p. 8

Implementation Details

• The implementation uses an ImageNet-pretrained ResNet-50, four Nvidia 1080 Ti GPUs, and a total batch size of 8.Source: Implementation Details · p. 8
• Training uses SGD with learning rate 0.01, weight decay 1e-4, momentum 0.9, λu = 1.0, and τ = 0.7.The localization branch uses K = 30; dynamic thresholding and reweighting use γ1 = 0.05 and γ2 = 0.6.Source: Implementation Details · p. 8
• For single-scale training, images are resized to short sides of 600 for PASCAL VOC and 800 for COCO; multi-scale training samples short sides from (640, 800).The long side is capped at 1,333 after resizing.Source: Implementation Details · p. 8