Distilling Object Detectors via Decoupled Features (2103.14475v1)

Abstract: 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. Our implementation is available at https://github.com/ggjy/DeFeat.pytorch.

• The paper presents a novel distillation approach that decouples neck and proposal features to capture essential information from both object and background regions.
• It employs feature decoupling using binary masks and separates positive/negative proposals to optimize the knowledge transfer process.
• Empirical results show notable gains, with a ResNet50-based Faster R-CNN improving from 37.4% to 40.9% mAP, validating the method's effectiveness.

Overview of "Distilling Object Detectors via Decoupled Features"

The research paper titled "Distilling Object Detectors via Decoupled Features" introduces an innovative approach to knowledge distillation for object detection. Unlike traditional distillation methodologies primarily focused on image classification, this work emphasizes the complexity of object detectors, which involve multiple intertwined loss functions where semantic information is not straightforwardly transferred. The authors argue that regions outside of object annotations, typically disregarded in distillation practices, hold valuable information crucial for enhancing student object detectors. Moreover, they contend that differentiating the importance of features from disparate regions is pivotal during knowledge distillation.

Core Contributions

1. Novel Distillation Approach: The authors propose a distillation strategy called decoupled features (DeFeat), distinguishing itself by processing two levels of decoupled features: decoupled neck features and decoupled proposals from the classification head. This division allows the embedding of critical information from both object and non-object regions into the student detector.
2. Empirical Validation: Extensive experimentation across diverse types of detectors with varying backbone architectures on standard benchmarks such as COCO demonstrates that DeFeat consistently outperforms state-of-the-art methods. For instance, a ResNet50-based Faster R-CNN student improved from 37.4% to 40.9% mAP, and a ResNet50-based RetinaNet increased from 36.5% to 39.7% mAP.

Key Insights

• Importance of Non-Object Regions: Contrary to mainstream approaches which prioritize object-centric regions, the paper highlights the utility of background or non-object regions in improving the student's performance. Distillation solely focusing on non-object regions reported comparable enhancements as those aimed at object regions.
• Feature Decoupling: By generating a binary mask based on ground truth bounding boxes, the intermediate FPN features were decoupled into object and background components for targeted distillation. This method circumvents the suboptimal results encountered when feature maps are treated uniformly throughout distillation.
• Proposal Decoupling: The paper also innovatively manages region proposals in the classification head by decoupling them into positive and negative subsets, optimizing the distillation process to reflect their inherent contributions to detection performance.

Implications and Future Directions

Practically, this research offers a significant stride towards more efficient object detectors without redesigning underlying architectures. Theoretical implications suggest a shift in how non-object information is evaluated in the context of machine learning tasks like object detection. Future research could delve into refining decoupling strategies further and exploring automatic optimization of the loss coefficients and temperature parameters used for different region types during distillation.

In conclusion, this work broadens the understanding of effective knowledge distillation, presenting promising avenues for future exploration in AI model compression and efficiency, especially pertinent as object detection systems continue to permeate real-world applications.