Source-linked AI summary
Self-Training and Adversarial Background Regularization for Unsupervised Domain Adaptive One-Stage Object Detection
Seunghyeon Kim, Jaehoon Choi, Taekyung Kim, Changick Kim
TL;DR
Different train and test distributions degrade supervised detectors, while unlabeled target data make reliable adaptation difficult. The paper introduces WST and BSR for one-stage detection, and reports improved performance under unsupervised domain adaptation.
Problem
Supervised detectors assume matching train and test distributions, but unsupervised adaptation must handle target-domain data without target labels.
Method
The paper combines weak self-training to reduce inaccurate pseudo-label effects with adversarial background score regularization to extract discriminative target-background features.
Results
The approach improves one-stage object detection performance under unsupervised domain adaptation.
Takeaways & Limitations
WST and BSR are complementary: BSR separates foregrounds from backgrounds, while WST strengthens class-wise discrimination.
Takeaways & Limitations
The setting assumes source and target distributions differ and provides no access to target labels.
Abstract
from arXiv · showhide
Deep learning-based object detectors have shown remarkable improvements. However, supervised learning-based methods perform poorly when the train data and the test data have different distributions. To address the issue, domain adaptation transfers knowledge from the label-sufficient domain (source domain) to the label-scarce domain (target domain). Self-training is one of the powerful ways to achieve domain adaptation since it helps class-wise domain adaptation. Unfortunately, a naive approach that utilizes pseudo-labels as ground-truth degenerates the performance due to incorrect pseudo-labels. In this paper, we introduce a weak self-training (WST) method and adversarial background score regularization (BSR) for domain adaptive one-stage object detection. WST diminishes the adverse effects of inaccurate pseudo-labels to stabilize the learning procedure. BSR helps the network extract discriminative features for target backgrounds to reduce the domain shift. Two components are complementary to each other as BSR enhances discrimination between foregrounds and backgrounds, whereas WST strengthen class-wise discrimination. Experimental results show that our approach effectively improves the performance of the one-stage object detection in unsupervised domain adaptation setting.
1. Introduction
The paper targets unsupervised domain adaptive one-stage detection, where train and test distributions differ and target labels are unavailable. It proposes WST and BSR to stabilize self-training and reduce domain shifts, improving target-domain detection.
- Unsupervised domain adaptation transfers knowledge from labeled source data to unlabeled target data when train and test distributions differ.
- Prior domain-adaptive detection methods either align non-transferable backgrounds, target two-stage detectors, or assume target image-level labels.
- Weak self-training (WST) reduces the negative effects of inaccurate pseudo-labels, including false positives and false negatives, without target labels.
- Adversarial background score regularization (BSR) reduces domain shifts by extracting discriminative features for target backgrounds, which is important for one-stage detectors without region proposals.
- The experiments show that the approach improves one-stage object detection under unsupervised domain adaptation.
2. Related Work
Related work covers detector architectures, domain adaptation, self-training, and prior domain-adaptive detection methods. The paper distinguishes its setting by combining one-stage detection with unsupervised target data and no image-level labels.
- Object Detection: Two-stage detectors use region proposals before classification and regression, whereas one-stage detectors directly estimate object coordinates and classes.
- Object Detection: One-stage detectors offer efficient inference but face foreground-background class imbalance, motivating methods such as focal loss.
- Domain Adaptation: Domain adaptation research uses distribution matching and adversarial learning to extract features that are more invariant across source and target domains.
- Domain Adaptation: Semantic-segmentation methods extend adaptation from feature-level alignment to pixel-level and feature-level distribution alignment.
- Self-Training: Self-training has been used to compensate for missing categorical information in classification and semantic segmentation.
- Domain-Adaptive Object Detection: Earlier object-detection adaptation methods target two-stage detectors, while one-stage methods may assume target image-level labels or use weakly supervised settings.
- Domain-Adaptive Object Detection: This paper instead develops WST for stable self-training without image-level labels and BSR for target-background discrimination rather than background alignment.
3. Proposed Method
The method combines weak self-training to reduce errors from inaccurate pseudo-labels with adversarial background score regularization to learn discriminative target-background features. Together, WST supplies category information while BSR reduces domain gaps and complements class-wise learning.
- Weak Self-Training: WST omits unreliable pseudo-label examples to reduce the effects of false positives and false negatives caused by domain shift.Reliable detections are selected by thresholding scores, while training is modified to limit harmful pseudo-label effects.
- Weak Self-Training: Weak negative mining selects ||Neg||/3 negative examples with the lowest confidence loss, avoiding hard negatives likely to be false negatives.Hard negative mining can select missed objects as background examples, such as the boat example in Fig. 3.
- Weak Self-Training: Supporting Region-based Reliable Score aggregates confidence from regions overlapping a final detection instead of relying on one detection score.Supporting regions are examples whose IoU with the final detection exceeds threshold δ, and SRRS combines IoU with class probability.
- Adversarial Background Score Regularization: BSR trains the classifier toward target background score t while training the feature extractor adversarially, encouraging discriminative features that deceive the classifier.The adversarial loss is applied to target inputs and implemented with a gradient reversal layer after relu4_3 of SSD300.
- Adversarial Background Score Regularization: BSR selects low-background-score examples rather than applying adversarial loss to all detector outputs, choosing low 3N examples batch-wise and adding a focal term.N is the number of examples predicted as foregrounds; t is set to 0.5 in the experiments.
- Combination: BSR and WST are complementary: BSR reduces background-related domain gaps, while WST provides category information through self-training.
4. Experiments
Experiments evaluate SSD300 for unsupervised adaptation from Pascal VOC to Clipart1k, Watercolor2k, and Comic2k using mAP. The proposed method improves performance, with gains varying by target domain and complementary BSR and WST components.
- Evaluation: Methods are evaluated with mean average precision (mAP), using confidence threshold 0.05 and IoU threshold 0.5.
- Methods: The comparison includes the SSD300 base network, DANN, naive self-training (ST), and the proposed method, with pseudo-labels recreated every iteration for ST and the proposed method.PL generates pseudo-labels once before training, whereas ST and the proposed method recreate them at every iteration.
- Results: On Clipart1k, BSR creates performance gaps of 8% and 2.2% mAP over the baseline and DANN, while combining BSR and WST further enhances performance.The improvement occurs without an additional discriminator network.
- Results: Self-training performs poorly for sheep, whereas DANN and BSR improve that class by nearly 9% AP.
- Results: On Comic2k, the proposed methods improve accuracy by about 5% mAP from the base, while DANN shows no improvement.On Watercolor2k, the algorithms show less improvement than on Clipart1k and Comic2k because images often contain single instances and lack hard backgrounds.
5. Analysis
Ablation and sensitivity analyses examine how WST components and BSR parameters affect stability and accuracy. Weak negative mining is especially important, while BSR is most effective when combined with WST components and appropriately tuned parameters.
- Study design: The ablation study evaluates WST components on Clipart1k, while parameter sensitivity experiments vary BSR's γ and t.
- WST ablation: Naive and partial self-training methods rapidly lose accuracy during training, indicating that SRRS or Mask is needed to stabilize self-training.The analysis links this behavior to false positives and false negatives in pseudo-labels.
- WST ablation: Weak Mask improves mAP by 14.5% without SRRS and 4.6% with SRRS compared with Mask, showing the value of weak negative mining.The authors report that selecting reliable negative samples is more effective than selecting positive samples in this setting.
- WST ablation: SRRS is ineffective alone but improves learning stability and accuracy when combined with Mask or Weak Mask.The experiments also indicate that reducing false positives remains important despite potentially reducing some true positives.
- BSR sensitivity: Smaller γ values impose stronger background regularization, whereas larger values relax regularization and ignore examples with background probability near 0.5.
- BSR sensitivity: t = 0.5 performs better than other tested values; alternative values cause overregularization and rapidly declining performance after learning-rate decay.
- Qualitative results: The proposed method detects some objects with lower confidence but greater correctness than the base network, including a chair missed by the base network.The authors attribute this qualitative behavior to BSR.
6. Conclusion
The paper addresses unsupervised domain adaptation for one-stage object detection by combining WST and BSR. WST handles inaccurate pseudo-labels, while BSR targets domain shifts in background features.
- Contributions: WST reduces false positives and false negatives by masking gradients of hard negative examples and using SRRS to select pseudo-labels.
- Contributions: BSR reduces domain shifts by extracting discriminative features for target backgrounds rather than aligning non-transferable features.
- Conclusion: The proposed approach improves one-stage object detection under unsupervised domain adaptation.