[论文解读] Towards End-to-end Semi-supervised Learning for One-stage Object Detection

Gen Luo, Yiyi Zhou+3 more|arXiv (Cornell University)|Feb 22, 2023

Advanced Neural Network Applications被引用 4

一句话总结

本文提出 OneTeacher，一种新颖的教师-学生框架，用于一阶段检测器（如 YOLOv5）的端到端半监督目标检测。通过多视角伪标签优化（MPR）和解耦半监督优化（DSO），解决了低质量伪标签和多任务优化冲突问题，在 COCO 上相比 Unbiased Teacher 实现了 15.0% 的相对 AP 提升。

ABSTRACT

Semi-supervised object detection (SSOD) is a research hot spot in computer vision, which can greatly reduce the requirement for expensive bounding-box annotations. Despite great success, existing progress mainly focuses on two-stage detection networks like FasterRCNN, while the research on one-stage detectors is often ignored. In this paper, we focus on the semi-supervised learning for the advanced and popular one-stage detection network YOLOv5. Compared with Faster-RCNN, the implementation of YOLOv5 is much more complex, and the various training techniques used in YOLOv5 can also reduce the benefit of SSOD. In addition to this challenge, we also reveal two key issues in one-stage SSOD, which are low-quality pseudo-labeling and multi-task optimization conflict, respectively. To address these issues, we propose a novel teacher-student learning recipe called OneTeacher with two innovative designs, namely Multi-view Pseudo-label Refinement (MPR) and Decoupled Semi-supervised Optimization (DSO). In particular, MPR improves the quality of pseudo-labels via augmented-view refinement and global-view filtering, and DSO handles the joint optimization conflicts via structure tweaks and task-specific pseudo-labeling. In addition, we also carefully revise the implementation of YOLOv5 to maximize the benefits of SSOD, which is also shared with the existing SSOD methods for fair comparison. To validate OneTeacher, we conduct extensive experiments on COCO and Pascal VOC. The extensive experiments show that OneTeacher can not only achieve superior performance than the compared methods, e.g., 15.0% relative AP gains over Unbiased Teacher, but also well handle the key issues in one-stage SSOD. Our source code is available at: https://github.com/luogen1996/OneTeacher.

研究动机与目标

• 解决一阶段检测器（如 YOLOv5）在半监督目标检测中被广泛使用但常被现有 SSOD 研究忽视的未充分探索挑战。
• 识别并解决一阶段 SSOD 中的两个关键问题：由于密集预测导致的低质量伪标签，以及共享标签标准引发的多任务优化冲突。
• 通过仔细调整 YOLOv5 的实现以契合 SSOD 最佳实践，最大化教师-学生学习的优势。
• 提出一种新的训练方案，使一阶段检测器能够实现端到端半监督学习，同时提升性能与鲁棒性。

提出的方法

• 引入多视角伪标签优化（MPR），通过增强视图对比与全局视图过滤提升伪标签质量，减少噪声预测。
• 通过修改网络结构实现解耦半监督优化（DSO），将分类头与回归头解耦，实现不同置信度阈值的任务特定伪标签生成。
• 通过原始视图与增强视图之间的实例级一致性检查，实现伪标签的跨视图精细化，提升其鲁棒性。
• 利用图像级预测结果过滤低置信度伪标签，减少误检，提升整体标签质量。
• 通过引入先进数据增强技术与优化策略（如余弦衰减、EMA）改进 YOLOv5 的训练流程，以最大化 SSOD 性能。
• 为分类任务与回归任务分别设置独立的伪标签阈值（均为 0.4），反映一阶段检测器在噪声容忍度上的差异。

Figure 1 : Qualities of pseudo-labels of YOLOv5 [ 22 ] and FasterRCNN [ 7 ] in Unbiased Teacher [ 11 ] . The subplots of (a) and (b) show that the average pseudo-label quality of YOLOv5 is much lower than that of FasterRCNN, especially in the initial training phase. Subplot (c) illustrates the diffe

实验结果

研究问题

• RQ1 为何现有教师-学生方法在 YOLOv5 等一阶段检测器上的半监督目标检测性能，相较于两阶段模型表现较差？
• RQ2 在因数千个锚点密集预测而产生高噪声的一阶段检测器中，如何提升伪标签质量？
• RQ3 分类与回归任务之间的多任务优化冲突在多大程度上损害了一阶段 SSOD 的性能？又该如何缓解？
• RQ4 在半监督设置下，YOLOv5 的最优数据增强方案与伪标签阈值配置为何？
• RQ5 能否设计一种统一的训练方案，在不破坏一阶段检测器固有架构的前提下，最大化半监督学习的优势？

主要发现

• 在标准 1× 设置下，OneTeacher 在 COCO 上相比 Unbiased Teacher 实现了 15.0% 的相对 AP 提升，证明其在一阶段 SSOD 中具有卓越性能。
• 所提出的 MPR 机制显著提升了伪标签质量，可视化结果表明误检减少，且先前遗漏的物体检测能力得到改善。
• DSO 通过解耦分类与回归头，有效缓解了多任务优化冲突，带来更快收敛速度与更优性能。
• 采用 YOLOv5 特定的数据增强策略而非标准两阶段 SSOD 方法，mAP 提升 13.5 个百分点，凸显了模型特异性实现的重要性。
• YOLOv5 的最优伪标签阈值为分类与回归任务均设为 0.4，显著低于 FasterRCNN 中使用的 0.7，表明一阶段检测器对噪声更敏感。
• 消融实验表明，MPR 中的增强视图精细化优于 Flip Ensemble，避免了通过未过滤平均导致的噪声放大。

Figure 2 : The framework of OneTeacher. The teacher network produces high-quality pseudo-labels for the student by the proposed Multi-view Pseudo-label Refinement (MPR). Its parameters are updated from the student ones via EMA [ 14 ] . OneTeacher also deploys a novel Decoupled-Semi-supervised Optimi