Introduction
3D object detection is a cornerstone of autonomous driving, yet current voxel-based frameworks often struggle with three persistent challenges: slow training convergence, performance degradation under heavy occlusion, and rigid pooling architectures that lack task-specific flexibility.
HPC-Net (Hybrid Pooling and Convolutional Network) addresses these hurdles by rethinking the fundamental building blocks of feature extraction. By introducing a modular pooling interface and specialized convolutions, HPC-Net achieves state-of-the-art accuracy while accelerating training convergence by over 44%.
The HPC-Net Architecture
The framework is designed as a high-precision voxel-based detector that integrates three novel components: Replaceable Pooling (RP), Depth Accelerated Convergence Convolution (DACConv), and the Multi-Scale Extended Receptive Field Feature Extraction Module (MEFEM).
1. Replaceable Pooling (RP)
Traditional 3D detectors often rely on fixed max or average pooling, which can limit generalization. The RP module provides a flexible interface using a three-step process: dimensionality expansion, pooling, and dimensionality reduction. This structure allows the network to adapt its pooling strategy to specific sensor data requirements, resulting in a 52.4% boost in efficiency compared to conventional 3D pooling approaches.
2. Depth Accelerated Convergence Convolution (DACConv)
To combat the notoriously slow training cycles of two-stage detectors, the authors introduced DACConv. Inspired by depthwise separable convolutions, DACConv employs an "over-parameterization" strategy during training. It splits the learning process between individual feature maps and input channels, allowing the model to capture complex features significantly faster. Crucially, this added complexity is only present during training; the module is simplified during inference to maintain real-time performance.
3. Multi-Scale Extended Receptive Field Feature Extraction Module (MEFEM)
Handling occluded or truncated objects—the "Hard" cases in datasets like KITTI—requires a more dynamic receptive field. MEFEM utilizes EAConv (Expanding Area Convolution), which leverages multi-layer Deformable Convolutions (DConv). Unlike standard fixed-grid convolutions, DConv can adapt its shape to match the geometry of a vehicle. By fusing these features across multiple scales, the network can infer the full shape of a car even when only small fragments are visible.
Experimental Results
HPC-Net was rigorously evaluated on the KITTI and Waymo Open Datasets. The results demonstrate significant improvements in both accuracy and training efficiency.
As shown in the table above, HPC-Net achieved the #1 position on the KITTI Car 2D Detection Ranking at the time of submission. More importantly, it excelled in the most challenging scenarios, ranking 1st in the "Hard" mode for 3D Car Detection with an 82.65% Average Precision (AP).
Beyond accuracy, the architectural innovations led to a 44.4% increase in training convergence speed. While traditional models might require 27 or more training cycles to reach peak performance, HPC-Net achieves comparable or superior results in approximately 15 cycles.
Conclusion
HPC-Net represents a significant step forward in making 3D object detection more practical for real-world deployment. By treating training speed and pooling flexibility as primary design goals alongside accuracy, the authors have created a framework that is both high-performing and efficient. Future work will explore extending these specialized convolutional strategies to smaller, non-rigid object classes such as pedestrians and cyclists, further enhancing the safety and reliability of autonomous systems.
Summary updated