MDPruner: Meta-Learning Driven Dynamic Filter Pruning for Efficient Object Detection

Filter pruning is a potent technique for diminishing the computational demands of Convolutional Neural Networks (CNNs), while effectively retaining model performance in image categorization tasks. However, research on its application to object detection models is scant and often incurs substantial performance degradation. In this paper, we present a novel Meta-network-based Dynamic Pruner method (MDPruner) designed to attenuate the computational burden of object detection models, concurrently preserving their efficacy. The rationale behind our approach is straightforward: for intricate images, we designate a lesser pruning rate to uphold detection accuracy, while for less complicated images, we apply a higher pruning rate to compress the model effectually and economize computational assets - thus, striking an ideal equilibrium between compression proficiency and performance sustenance. To accomplish this, we exploit a supernet framework to concurrently train pruned networks of diverse sizes, thereby obviating the necessity for supplementary parameter overhead. Furthermore, to deftly and accurately evaluate the detection complexity of input images and to choose the optimal pruning network, we propose a lightweight meta-network module, which is engineered to identify the most suitable subnetwork for processing, dependent on the image's detection complexity. Comprehensive evaluations effectuated on the COCO and PASCAL VOC datasets have substantiated the effectiveness of our proposed MDPruner in compressing object detection models.