Simple Supervision Is Hard to Beat: A Bitter Lesson from Sparse Target Labels in Domain-Adaptive Object Detection
RTSM method improves sparse-label domain-adaptive object detection by 1.7 to 18.3 AP50.
Key Findings
Methodology
The paper introduces Random-Target Supervised Mixing (RTSM), which incorporates sparse target annotations through a supervised detection loss, enhancing domain-adaptive object detection performance. RTSM maintains the original unlabeled adaptation branch unchanged, providing a simple yet effective anchor.
Key Results
- RTSM improved AP50 by 1.7 to 18.3 across four SFDA-OD methods. For instance, adding 5% randomly labeled target data to DDT increased AP50 from 53.669 to 58.695.
- RTSM consistently improved performance across different target-label budgets.
- Ten feedback plugins were evaluated, showing limited and method-dependent gains.
Significance
The study reveals that simple supervision is hard to beat in sparse-label SFDA-OD. RTSM provides a simple yet effective anchor, emphasizing the importance of direct supervision.
Technical Contribution
RTSM significantly enhances domain-adaptive object detection performance through direct supervision of sparse target labels, offering a simple yet strong baseline without complex label feedback mechanisms.
Novelty
RTSM is the first to demonstrate the powerful effect of direct supervision in sparse-label SFDA-OD, showing more stable performance improvements compared to complex label feedback mechanisms.
Limitations
- RTSM does not fully utilize sparse labels to improve the unlabeled self-training branch.
- Feedback plugin gains are limited and method-dependent.
Future Work
Future research could explore more effective ways to utilize sparse labels to improve the unlabeled self-training branch or develop new label feedback mechanisms.
AI Executive Summary
In domain-adaptive object detection, existing methods typically rely on teacher-student self-training frameworks using pseudo-labels to adapt to unlabeled target domains. However, these methods perform limitedly in sparse-label scenarios. The paper introduces Random-Target Supervised Mixing (RTSM), which enhances domain-adaptive object detection performance by incorporating supervised detection loss on sparsely annotated target images. RTSM consistently improves performance across four SFDA-OD methods, revealing the powerful effect of simple supervision in sparse-label scenarios. Despite attempts with various label feedback plugins, gains were limited and method-dependent, emphasizing the importance of direct supervision. Future research could explore more effective ways to utilize sparse labels to improve the unlabeled self-training branch or develop new label feedback mechanisms.
Deep Analysis
Background
Domain-adaptive object detection aims to address the performance degradation of object detectors in environments with different visual statistics. Existing methods typically rely on teacher-student self-training frameworks using pseudo-labels to adapt to unlabeled target domains. However, these methods perform limitedly in sparse-label scenarios.
Core Problem
In sparse-label scenarios, existing pseudo-label-dependent methods may produce false positives, miss hard target objects, and provide unstable optimization signals. Effectively utilizing sparse labels to improve domain-adaptive object detection is an important and challenging problem.
Innovation
RTSM enhances domain-adaptive object detection performance by incorporating supervised detection loss on sparsely annotated target images. Compared to complex label feedback mechanisms, RTSM provides a simple yet strong baseline.
Methodology
- �� RTSM significantly enhances domain-adaptive object detection performance through direct supervision of sparse target labels. • Maintains the original unlabeled adaptation branch unchanged, providing a simple yet effective anchor. • Evaluated ten feedback plugins, showing limited and method-dependent gains.
Experiments
Experiments were conducted on Cityscapes, Foggy Cityscapes, and BDD100K datasets, evaluating four SFDA-OD methods and two object detectors. RTSM consistently improved performance across different target-label budgets.
Results
RTSM improved AP50 by 1.7 to 18.3 across four SFDA-OD methods. For instance, adding 5% randomly labeled target data to DDT increased AP50 from 53.669 to 58.695.
Applications
RTSM can be applied in domain-adaptive object detection for autonomous driving and surveillance scenarios, providing a simple yet effective solution.
Limitations & Outlook
RTSM does not fully utilize sparse labels to improve the unlabeled self-training branch. Feedback plugin gains are limited and method-dependent.
Plain Language Accessible to non-experts
Imagine a factory where workers are on the production line. RTSM is like giving workers a simple tool to help them do their jobs better, rather than complex machinery. Although the factory has many different machines, this simple tool significantly boosts productivity.
ELI14 Explained like you're 14
Imagine you're playing a game with many levels. RTSM is like giving you a super power-up that makes it easier to pass levels, rather than complex strategies. Although the game has many different power-ups, this super power-up significantly boosts your gaming experience.
Glossary
RTSM (Random-Target Supervised Mixing)
A method that incorporates supervised detection loss on sparsely annotated target images.
Used to enhance domain-adaptive object detection performance.
SFDA-OD (Source-Free Domain Adaptive Object Detection)
A method to adapt source-trained detectors to unlabeled target domains.
Implemented through teacher-student self-training frameworks.
Pseudo-labels
Labels predicted by the teacher model used to supervise the student model.
Used in unlabeled target domains.
Teacher-Student Self-Training
A method where the teacher model predicts pseudo-labels and optimizes the student model.
Used in domain-adaptive object detection.
AP50
Average precision at an IoU threshold of 0.5.
Used to evaluate object detector performance.
Open Questions Unanswered questions from this research
- 1 How to more effectively utilize sparse labels to improve the unlabeled self-training branch remains an open question.
- 2 The design of label feedback mechanisms needs further research to enhance their gains.
Applications
Immediate Applications
Autonomous Driving
RTSM can be applied in domain-adaptive object detection for autonomous driving scenarios, providing a simple yet effective solution.
Long-term Vision
Intelligent Surveillance
RTSM can be applied in intelligent surveillance scenarios, helping to improve the accuracy and efficiency of surveillance systems.
Abstract
Source-free domain adaptive object detection adapts a source-trained detector to an unlabeled target domain, typically through teacher-student self-training with pseudo-labels. We revisit this setting when a small, uniformly sampled subset of target images is labeled. We introduce Random-Target Supervised Mixing (RTSM), a simple anchor that incorporates these annotations through a supervised detection loss while leaving the original unlabeled adaptation branch unchanged. Across evaluations spanning four SFDA-OD methods, two object detectors, multiple adaptation tasks, and target-label budgets from 1% to 10%, RTSM consistently improves pure SFDA by 1.7 to 18.3 AP50. We then examine whether the same annotations can provide further gains by steering unlabeled self-training. To this end, we evaluate ten sparse-label feedback plugins covering pseudo-label selection, object completion, and optimization control, which yield limited and method-dependent gains over RTSM. These results reveal a bitter lesson for sparse-label SFDA-OD: simple supervision is hard to beat. RTSM therefore provides a simple yet effective anchor for sparse-label SFDA-OD.