General Instance Distillation for Object Detection

Dai, Xing; Jiang, Zeren; Wu, Zhao; Bao, Yiping; Wang, Zhicheng; Liu, Si; Zhou, Erjin

doi:10.1109/cvpr46437.2021.00775

Cited by 184 publications

(100 citation statements)

References 19 publications

Supporting

Mentioning

100

Contrasting

Order By: Relevance

“…For example, under the guidance of RseNext101 teacher detector, the ResNet50 with RetinaNet student detector achieves 40.1% in mAP, surpassing [32] by 0.5% mAP. With the help of ResNet101 teacher detector, the ResNet50 based RetinaNet gets 39.3% mAP, which also exceeds the State-of-the-art methods General Instance [5]. The results demonstrate that our method has achieved a good performance improvement, surpassing the previous SOTA methods.…”

Section: Comparison With State-of-the-artsmentioning

confidence: 82%

“…Ruoyu et al [26] propose to use a two-dimensional Gaussian mask to suppress irrelevant redundant background while emphasizing the target information. Dai et al [5] propose a distillation method for detection tasks based on discriminative instances without considering the positive or negative distinguished by ground truth. Guo et al [8] point out that the information of features derived from regions excluding objects is also essential for distilling the student detector, which did not explore the background area in more detail.…”

Section: Knowledge Distillationmentioning

confidence: 99%

See 1 more Smart Citation

Distilling Object Detectors with Feature Richness

Du¹,

Zhang²,

Chang³

et al. 2021

Preprint

View full text Add to dashboard Cite

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, Reti-naNet 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%.

show abstract

Section: Comparison With State-of-the-artsmentioning

confidence: 82%

Section: Knowledge Distillationmentioning

confidence: 99%

Distilling Object Detectors with Feature Richness

Du¹,

Zhang²,

Chang³

et al. 2021

Preprint

View full text Add to dashboard Cite

show abstract

“…This trend was initiated by Chen et al [5], which proposed to distill knowledge from a teacher detector to a student detector in both the backbone and head stages. Then, Wang et al [38] proposed to restrict the teacher-student feature imitation to regions around positive anchor boxes; Dai et al [8] produced general instances based on both the teacher's and student's outputs, and distilled feature-based, relation-based and response-based knowledge in these general instances; Guo et al [12] proposed to decouple the intermediate features and classification predictions of the positive and negative regions during knowledge distillation. All the aforementioned knowledge distillation methods require the student and the teacher to follow the same kind of detection framework, and thus typically transfer knowledge between models that only differ in terms of backbone, such as from a RetinaNet-ResNet152 to a RetinaNet-ResNet50.…”

Section: Related Workmentioning

confidence: 99%

“…While, as will be shown by our experiments, knowledge distillation for classification already helps the student detector, it does not aim to improve its localization performance. Nevertheless, localization, or bounding box regression, is critical for the success of a detector and is typically addressed by existing detector-to-detector distillation frameworks [5,8]. To also tackle this in our classifier-to-detector approach, we develop a feature-level distillation strategy, exploiting the intuition that the intermediate features extracted by the classification teacher from a bounding box produced by the student should match those of the ground-truth bounding box.…”

Section: Kd Loc : Knowledge Distillation For Localizationmentioning

confidence: 99%

“…We then compare our classifier-to-detector distillation approach with the state-of-the-art detector-todetector ones, such as KD [5], FGFI [38], GID [8] and FKD [42]. Here, in addition to the compact students used in Section 4.1, we also report results on the larger students that are commonly used in the literature, i.e., Faster RCNN and RetinaNet with deeper ResNet50 (R50) backbones.…”

Section: Comparison With Detector-to-detector Distillationmentioning

confidence: 99%

See 1 more Smart Citation

Distilling Image Classifiers in Object Detectors

Guo

Álvarez

Salzmann

2021

Preprint

View full text Add to dashboard Cite

Knowledge distillation constitutes a simple yet effective way to improve the performance of a compact student network by exploiting the knowledge of a more powerful teacher. Nevertheless, the knowledge distillation literature remains limited to the scenario where the student and the teacher tackle the same task. Here, we investigate the problem of transferring knowledge not only across architectures but also across tasks. To this end, we study the case of object detection and, instead of following the standard detector-to-detector distillation approach, introduce a classifier-to-detector knowledge transfer framework. In particular, we propose strategies to exploit the classification teacher to improve both the detector's recognition accuracy and localization performance. Our experiments on several detectors with different backbones demonstrate the effectiveness of our approach, allowing us to outperform the state-of-the-art detector-to-detector distillation methods.

show abstract