[论文解读] Distilling Object Detectors with Feature Richness

Zhixing Du, Rui Zhang+5 more|arXiv (Cornell University)|Nov 1, 2021

Advanced Neural Network Applications参考文献 31被引用 36

一句话总结

本文提出 Feature Richness Score (FRS) 用于在蒸馏对象检测器时选择信息丰富的特征，利用盒外的 bounding boxes 并修剪盒内误分类特征，在基于锚点、无锚点以及两阶段检测器上的表现均有提升。

ABSTRACT

In recent years, large-scale deep models have achieved great success, but the huge computational complexity and massive storage requirements make it a great challenge to deploy them in resource-limited devices. As a model compression and acceleration method, knowledge distillation effectively improves the performance of small models by transferring the dark knowledge from the teacher detector. However, most of the existing distillation-based detection methods mainly imitating features near bounding boxes, which suffer from two limitations. First, they ignore the beneficial features outside the bounding boxes. Second, these methods imitate some features which are mistakenly regarded as the background by the teacher detector. To address the above issues, we propose a novel Feature-Richness Score (FRS) method to choose important features that improve generalized detectability during distilling. The proposed method effectively retrieves the important features outside the bounding boxes and removes the detrimental features within the bounding boxes. Extensive experiments show that our methods achieve excellent performance on both anchor-based and anchor-free detectors. For example, RetinaNet with ResNet-50 achieves 39.7% in mAP on the COCO2017 dataset, which even surpasses the ResNet-101 based teacher detector 38.9% by 0.8%. Our implementation is available at https://github.com/duzhixing/FRS.

研究动机与目标

• 将对象检测的知识蒸馏动机聚焦于超越边界框的信息丰富特征。
• 提出 Feature Richness Score (FRS)，用于识别跨越所有区域的对象样特征，而不仅限于框内。
• 展示一个可插拔的蒸馏框架，适用于基于锚点/无锚点以及两阶段检测器。
• 通过利用盒外信息并减少盒内误分类特征，证明 FRS 提高了通用检测能力。

提出的方法

• 将特征丰富性 S 定义为使用教师分类分数 P(c|f, theta) 在各类别中的最大值。
• 由相应的教师分类分数计算每个金字塔层的特征丰富性掩码 S_l。
• 使用掩码对像素级蒸馏损失进行加权，蒸馏 FPN 层（L_FPN）和分类头（L_head）。
• 将损失与标准 GT 损失结合：L = L_GT + alpha L_FPN + beta L_head.

实验结果

研究问题

• RQ1 特征丰富区域（盒外）是否能为对象检测蒸馏提供有益的指导？
• RQ2 通过像素级特征丰富性掩码对蒸馏进行加权是否能在不同检测器类型上提升学生模型的性能？
• RQ3 FRS 在 COCO 的基于锚点、无锚点和两阶段检测器上的表现如何？
• RQ4 盒外特征和盒内错误标注特征在多大程度上影响蒸馏质量？

主要发现

模式	mAP	AP50	AP75	AP_S	AP_M	AP_L
Retina-Res101(teacher) 2x	38.9	58.0	41.5	21.0	42.8	52.4
Retina-Res50(student) 2x	37.4	56.7	39.6	20.0	40.7	49.7
ours 2x	39.7	58.6	42.4	21.8	43.5	52.4

• FRS 在 RetinaNet、GFL、FCOS 和 Faster R-CNN 上均有提升，学生采用 ResNet-50，教师为 ResNet-101。
• RetinaNet-Res50 2x 调度下达到 39.7% mAP，超过教师的 38.9% 提升 0.8%。
• GFL-Res50 (1x) 相较基线提升 3.4% mAP；在 2x 设置下某些情况下增益超过 1.8–4.2% 的各指标。
• FCOS-Res50 (2x) 实现 40.9% mAP，超过教师的性能。
• 消融实验显示 FPN 和分类头蒸馏都贡献了增益，联合实现的提升最高可达 2.3% mAP。
• 定性与熵分析表明 TP+FP 区域（盒外信息丰富区域）对蒸馏尤其有益。