Weighted Copula Entropy for Structural Pruning in Long-Tailed Autonomous Driving Object Detection

TL;DR

This paper introduces a new structural pruning method for autonomous driving object detection that improves efficiency without sacrificing accuracy, especially for rare classes.

Contribution

The novel framework uses weighted copula entropy and Elastic Net regularization to prune deep networks while preserving rare-class discriminators.

Findings

At 50% pruning rate, FLOPs and parameters are reduced by nearly 50% with minimal mAP@0.5 loss.

The mAP of the extreme tail class Train improves from 0% to 3.84% and 2.76% to 5.12% after pruning.

The method achieves better accuracy-efficiency trade-offs than mainstream pruning approaches.

Abstract

In autonomous driving, deep convolutional neural networks face a core conflict between computational efficiency and safety-critical robustness on resource-constrained onboard computing units. Dominant structural pruning, based on weight magnitude or geometric statistics, fails in long-tailed traffic scenarios by equating parameter magnitude with feature importance and pruning critical filters in the tail classes. To address this, we propose a structural pruning framework that evaluates the semantic utility of features using weighted copula entropy rather than relying solely on their magnitude. Our novel approach integrates Elastic Net regularization for inducing sparsity and weighted copula entropy for unbiased information-theoretic feature selection. By incorporating inverse class frequency weighting into empirical Copula estimation, we decouple feature relevance from sample abundance, ensuring the preservation of rare-class discriminators based on their information content rather than occurrence frequency. Furthermore, this metric is embedded into an enhanced max-relevance and min-redundancy algorithm to eliminate semantic redundancy while maintaining representational diversity. Extensive experiments on the BDD100K dataset with YOLOv5l and YOLOv8l architectures demonstrate that, at a 50% pruning rate, the proposed method reduces FLOPs and parameters by nearly 50%, with only 0.09% mAP@0.5 loss for YOLOv5l and 0.14% mAP@0.5 loss for YOLOv8l, while significantly improving the mAP of the extreme tail class Train from 0% to 3.84% and 2.76% to 5.12%, respectively. It achieves a more favorable trade-off between detection accuracy and computational efficiency than mainstream pruning approaches. This work provides a lightweight scheme for autonomous driving perception models and a new information-theoretic perspective for structured network pruning.

Figures