[论文解读] Unbiased Teacher v2: Semi-supervised Object Detection for Anchor-free and Anchor-based Detectors

Yen‐Cheng Liu, Chih‐Yao Ma+1 more|arXiv (Cornell University)|Jun 19, 2022

Advanced Image and Video Retrieval Techniques被引用 4

一句话总结

本文提出 Unbiased Teacher v2，一种半监督目标检测框架，通过引入 Listen2Student——一种新颖的基于不确定性的伪标签选择机制，实现对无锚点（anchor-free）和有锚点（anchor-based）检测器的通用化。该方法通过比较教师网络与学生网络之间的相对不确定性，减少误导性伪标签，从而在 COCO-standard、COCO-additional 和 VOC 基准上取得最先进性能。

ABSTRACT

With the recent development of Semi-Supervised Object Detection (SS-OD) techniques, object detectors can be improved by using a limited amount of labeled data and abundant unlabeled data. However, there are still two challenges that are not addressed: (1) there is no prior SS-OD work on anchor-free detectors, and (2) prior works are ineffective when pseudo-labeling bounding box regression. In this paper, we present Unbiased Teacher v2, which shows the generalization of SS-OD method to anchor-free detectors and also introduces Listen2Student mechanism for the unsupervised regression loss. Specifically, we first present a study examining the effectiveness of existing SS-OD methods on anchor-free detectors and find that they achieve much lower performance improvements under the semi-supervised setting. We also observe that box selection with centerness and the localization-based labeling used in anchor-free detectors cannot work well under the semi-supervised setting. On the other hand, our Listen2Student mechanism explicitly prevents misleading pseudo-labels in the training of bounding box regression; we specifically develop a novel pseudo-labeling selection mechanism based on the Teacher and Student's relative uncertainties. This idea contributes to favorable improvement in the regression branch in the semi-supervised setting. Our method, which works for both anchor-free and anchor-based methods, consistently performs favorably against the state-of-the-art methods in VOC, COCO-standard, and COCO-additional.

研究动机与目标

• 为解决当前半监督目标检测（SS-OD）方法在无锚点检测器中缺乏适用性的问题，这些检测器在目标检测中日益流行。
• 解决在半监督设置下，尤其在使用置信度阈值时，边界框回归中伪标签化效果不佳的问题。
• 通过比置信度阈值更鲁棒的选择机制，减少误导性伪标签，从而提升回归性能。
• 通过应用统一且高效的方法，弥合无锚点与有锚点检测器在半监督学习下的性能差距。
• 通过在包括 COCO-standard、COCO-additional 和 VOC 在内的多样化基准上进行实证验证，证明所提方法的有效性。

提出的方法

• 提出 Listen2Student 机制，基于教师网络与学生网络之间的相对不确定性，选择回归伪标签。
• 对每个边界框的四个坐标分别进行边界级不确定性估计，以指导伪标签选择，而非依赖分类置信度。
• 仅在教师网络的不确定性低于学生网络时，才施加无监督回归损失，从而减少来自不可靠预测的噪声。
• 将该机制应用于无锚点（如 FCOS、RetinaNet）和有锚点（如 Faster R-CNN）检测器，确保在不同架构间的通用性。
• 采用一致的自训练框架，结合数据增强与一致性正则化，仅对高置信度、低不确定性区域应用伪标签。
• 采用多阶段训练策略，通过迭代伪标签生成与模型优化，保持教师与学生预测之间的一致性。

Figure 1 : To improve the unsupervised regression loss, we propose (a) Listen2Student , which explicitly compares the prediction uncertainties between the Teacher and the Student and selects these instances where the teacher has lower uncertainty than the student. We then enforce the unsupervised re

实验结果

研究问题

• RQ1 为何现有 SS-OD 方法在无锚点检测器上的表现显著劣于有锚点检测器？
• RQ2 在半监督设置下，中心度（centerness）与定位相关标签的可靠性如何下降？
• RQ3 教师与学生之间的相对不确定性能否提升边界框回归的伪标签质量？
• RQ4 基于边界的不确定性估计机制是否优于基于置信度的阈值方法，从而带来更好的回归性能？
• RQ5 统一的 SS-OD 框架能否在无锚点与有锚点检测器上均实现最先进性能？

主要发现

• 仅使用 COCO2017-train 作为标注数据、COCO2017-unlabeled 作为未标注数据，Unbiased Teacher v2 在 COCO-additional 上达到 44.75 mAP，优于监督基线的 40.90 mAP。
• 在 VOC 上，使用 VOC07 作为标注数据、VOC12+COCO20cls 作为未标注数据，模型达到 58.08 mAP，超过监督基线，展现出强大的泛化能力。
• Listen2Student 机制在所有 IoU 阈值（AP55 至 AP95）下均提升 mAP，尤其在严格指标 AP95 上增益最大，证实了边界定位精度的提升。
• 仅使用置信度阈值会降低 AP95 性能，表明其引入了有害的误导性伪标签，影响精确定位。
• 该方法显著缩小了无锚点与有锚点检测器在半监督学习下的性能差距，表明 SS-OD 可有效推广至现代无锚点架构。
• 消融实验表明，基于相对不确定性的选择机制在高精度检测场景下，显著优于基于置信度的选择机制。

Figure 2 : Illustration of Centerness bias issue. (a) Selecting pseudo-boxes based on box scores leads to worse results in semi-supervised learning compared with selecting based on classification scores. (b) Box scores of the anchor-free detectors [ 29 , 37 ] are defined as the multiplication of the