Distilling Object Detectors via Decoupled Features

By Jianyuan Guo, Kai Han, Yunhe Wang, Han Wu, Xinghao Chen, Chunjing Xu. 2021

258 citations recorded by the source corpus.

Knowledge distillation is a widely used paradigm for inheriting information from a complicated teacher network to a compact student network and maintaining the strong performance. Different from image classification, object detectors are much more sophisticated with multiple loss functions in which features that semantic information rely on are tangled. In this paper, we point out that the information of features derived from regions excluding objects are also essential for distilling the student detector, which is usually ignored in existing approaches. In addition, we elucidate that features from different regions should be assigned with different importance during distillation. To this end, we present a novel distillation algorithm via decoupled features (DeFeat) for learning a better student detector. Specifically, two levels of decoupled features will be processed for embedding useful information into the student, i.e., decoupled features from neck and decoupled proposals from classification head. Extensive experiments on various detectors with different backbones show that the proposed DeFeat is able to surpass the state-of-the-art distillation methods for object detection. For example, DeFeat improves ResNet50 based Faster R-CNN from 37.4% to 40.9% mAP, and improves ResNet50 based RetinaNet from 36.5% to 39.7% mAP on COCO benchmark. Code will be released1,2.

Subfields: Computer Vision and Pattern Recognition

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Computer Vision and Pattern Recognition

1. Kai Han
2. Yunhe Wang
3. Jianyuan Guo
4. Chuanjian Liu
5. S. Cong
6. Ying Nie
7. Yuan Gong
8. Chun Yuan

Citations in the corpus

• Overcoming Catastrophic Forgetting in Incremental Object Detection via Elastic Response Distillation (2022, positive): “[13] believing that foreground and background both play an unique role in object detection, proposed an object detection distillation method that decoupled foreground and background.”
• Gold-YOLO: Efficient Object Detector via Gather-and-Distribute Mechanism (2023, background): “[12, 7, 17] studied the optimal network structure through NAS for object detection task, and [16, 23, 19] explore another way to improve the performance of the model by distillation.”
• Focal and Global Knowledge Distillation for Detectors (2021, methodology): “[7] shows that both the foreground and background play important roles for distillation, and distilling them separately benefits the student more.”
• A review of convolutional neural network architectures and their optimizations (2022, background): “considered that the feature information obtained in regions excluding objects is usually ignored in the target detection task for student networks (Guo et al. 2021).”