[论文解读] Localization Distillation for Object Detection

Zhaohui Zheng, Rongguang Ye+5 more|arXiv (Cornell University)|Apr 12, 2022

Advanced Neural Network Applications被引用 14

一句话总结

本文提出定位蒸馏（Localization Distillation, LD），一种用于目标检测的新颖知识蒸馏方法，通过边界框分布传递定位知识，使logit模仿优于传统的特征模仿。通过引入有价值的定位区域（Valuable Localization Regions, VLR）实现选择性蒸馏，LD在不降低推理速度的前提下提升了检测精度，在MS COCO、PASCAL VOC和DOTA基准上实现了最先进（SOTA）的AP提升。

ABSTRACT

Previous knowledge distillation (KD) methods for object detection mostly focus on feature imitation instead of mimicking the prediction logits due to its inefficiency in distilling the localization information. In this paper, we investigate whether logit mimicking always lags behind feature imitation. Towards this goal, we first present a novel localization distillation (LD) method which can efficiently transfer the localization knowledge from the teacher to the student. Second, we introduce the concept of valuable localization region that can aid to selectively distill the classification and localization knowledge for a certain region. Combining these two new components, for the first time, we show that logit mimicking can outperform feature imitation and the absence of localization distillation is a critical reason for why logit mimicking underperforms for years. The thorough studies exhibit the great potential of logit mimicking that can significantly alleviate the localization ambiguity, learn robust feature representation, and ease the training difficulty in the early stage. We also provide the theoretical connection between the proposed LD and the classification KD, that they share the equivalent optimization effect. Our distillation scheme is simple as well as effective and can be easily applied to both dense horizontal object detectors and rotated object detectors. Extensive experiments on the MS COCO, PASCAL VOC, and DOTA benchmarks demonstrate that our method can achieve considerable AP improvement without any sacrifice on the inference speed. Our source code and pretrained models are publicly available at https://github.com/HikariTJU/LD.

研究动机与目标

• 挑战长期以来认为在目标检测知识蒸馏中特征模仿始终优于logit模仿的信念。
• 探究为何多年来logit模仿在目标检测中表现不佳，尤其是由于缺乏有效的定位知识迁移。
• 开发一种简单而有效的蒸馏框架，通过显式建模定位不确定性，使logit模仿能够超越特征模仿。
• 引入有价值的定位区域（VLR）概念，基于区域重要性选择性地蒸馏分类与定位知识。
• 建立LD与分类知识蒸馏（KD）之间的理论联系，表明二者在优化效果上具有等价性。

提出的方法

• 提出定位蒸馏（LD），利用教师网络的边界框分布作为监督信号，以传递比伪边界框回归更丰富的定位知识。
• 采用离散化概率分布建模定位模糊性，其中平坦度与锐度反映不确定性，从而实现更有效的定位知识迁移。
• 引入有价值的定位区域（VLR）概念，识别并聚焦于高重要性区域进行蒸馏，提升效率与性能。
• 在统一的logit模仿框架中结合LD与分类KD，实现分类与定位知识的独立蒸馏。
• 采用梯度重标度分析表明，LD在logits空间中隐式执行类似焦点损失（focal loss）的梯度重加权，提升训练稳定性和收敛性。
• 通过概率分布的线性组合分解教师输出，实现LD与分类KD在优化行为上的理论等价性。

Figure 1: Existing KD pipelines for object detection. ① Logit Mimicking: classification KD in [ 1 ] . ② Feature Imitation: recent popular methods distill intermediate features based on various distillation regions, which usually need adaptive layers to align the size of the student’s feature map. ③

实验结果

研究问题

• RQ1 为何尽管在图像分类中表现成功，logit模仿在目标检测中仍长期表现不如特征模仿？
• RQ2 如果能有效传递定位知识，logit模仿是否能在目标检测中超越特征模仿？
• RQ3 如何建模并利用定位模糊性以提升蒸馏性能？
• RQ4 选择性蒸馏在提升知识蒸馏效率与准确性方面起到何种作用？
• RQ5 在优化动力学方面，定位蒸馏与分类知识蒸馏是否存在理论等价性？

主要发现

• 定位蒸馏（LD）首次使logit模仿在目标检测中超越特征模仿，证明了缺乏定位蒸馏是logit模仿历史表现不佳的关键原因。
• 所提方法在MS COCO、PASCAL VOC和DOTA基准上实现了显著的AP提升，且未增加推理延迟，证明了其高效性与有效性。
• 引入有价值的定位区域（VLR）通过聚焦高重要性区域，显著提升了蒸馏性能，同时增强了准确率与训练稳定性。
• 理论分析表明，LD与分类KD具有相同的优化效果，验证了其在学习动态上的等价性，提供了坚实的理论基础。
• 梯度重标度分析揭示，LD隐式执行类似焦点损失的梯度重加权，有助于缓解训练初期的困难并改善收敛性。
• 大量消融实验表明，LD显著降低了定位模糊性，学习到更鲁棒的特征表示，并增强了学生检测器的泛化能力。

Figure 2: Bottom edge for “ elephant ” and right edge for “ surfboard ” are ambiguous to locate.