AutoPKD: automatic channel pruning and patch-based relation distillation for efficient 3D object detection

Wenxiang Li

doi:10.1016/j.eswa.2026.133994

Return

AutoPKD: automatic channel pruning and patch-based relation distillation for efficient 3D object detection

2026-08-22

PRE

DOI：10.1016/j.eswa.2026.133994

Original

Original request for help

Save

Abstract

En 中文

• Automatic pruning and relation distillation for 3D object detection. • ABC-based channel search for structured model pruning. • Training-free proxy metric for rapid structure evaluation. • Real-world deployment on autonomous vehicle platform. Abstract LiDAR-based 3D object detection has achieved remarkable progress, but over-parameterized models hinder deployment in resource-limited scenarios. While channel pruning and knowledge distillation are well-studied in 2D vision, their applications in 3D object detection remain underexplored. In addition, most pruning methods for 3D object detection are rule-of-thumb, often resulting in suboptimal structures. To tackle this, we propose AutoPKD, a joint model compression framework that integrates Automatic Channel Search Pruning (ACSP) and Patch-based Relation Knowledge Distillation (PRKD). ACSP, based on the Artificial Bee Colony (ABC) algorithm, automatically searches for the optimal pruned structure. However, evaluating pruned structures in the standard ABC requires retraining, which incurs substantial computational overhead. To address this, we propose a training-free Category-Aware Efficient Fitness Evaluation (CaEFE) method, which evaluates a pruned structure in around 20 seconds on KITTI dataset, whereas the standard training-based method may take minutes or even hours and incurs substantial computational overhead. To mitigate the accuracy loss from pruning, PRKD captures both inter-patch and intra-patch scene-level information in Bird’s Eye View (BEV), enabling the student model to learn structural relations rather than directly aligning feature values. Extensive experiments show that AutoPKD effectively compresses 3D detection models while maintaining competitive accuracy. In addition, efficiency evaluations on the NVIDIA Jetson AGX Orin (32GB) platform provide practical evidence that deployment performance depends on both model compression and hardware characteristics, demonstrating the potential of AutoPKD for resource-constrained scenarios.

Keywords:

Channel pruning

Knowledge distillation

Automatic structure search

3D object detection