Abstract
Single-stage point-based 3D object detectors are efficient but often suffer from inadequate learning of low-quality objects (ILQ) and misalignment between localization accuracy and classification confidence (MLC). We propose SGCCNet to address these issues. SGCCNet utilizes a Saliency-Guided Data Augmentation (SGDA) strategy to reduce reliance on salient features and a Confidence Correction Mechanism (CCM) to align classification scores with localization quality. On the KITTI benchmark, SGCCNet achieves 80.82% AP3D on the Moderate level, outperforming existing point-based detectors while maintaining real-time inference speeds.
Introduction
LiDAR-based 3D object detection is critical for autonomous vehicle perception. While single-stage point-based detectors offer lightweight and fast inference, they face two major hurdles. First, Insufficient Learning of Low-Quality Objects (ILQ) occurs because neural networks tend to focus on the most prominent (salient) features, making them fragile when these features are missing. Second, Misalignment between Localization and Classification (MLC) leads to high-confidence scores for poorly localized boxes, which degrades performance during post-processing. SGCCNet is designed to tackle these challenges directly.
Methodology
The SGCCNet architecture integrates novel data augmentation and post-processing techniques into a point-based detection framework.
Saliency-Guided Data Augmentation (SGDA)
To address ILQ, SGDA identifies salient points by calculating how much shifting a point toward the object's centroid affects classification loss. During training, these high-saliency points are intentionally dropped, forcing the model to learn from more subtle, low-quality features. This "differentiable erasure" process makes the model significantly more robust to occlusions and sparse data.
Backbone Enhancements
Dropping points can introduce instability. To counter this, SGCCNet includes a Geometric Normalization Module (GNM)—a specialized normalization for local point groups—and Skip Connection Blocks (SCB) to preserve contextual features during downsampling.
Confidence Correction Mechanism (CCM)
CCM addresses the MLC problem by refining classification scores in the post-processing stage. Instead of relying on a single point's prediction, CCM aggregates information from neighboring key points. If multiple neighbors support a high-quality bounding box, the confidence is boosted, ensuring that the best-aligned boxes are preserved during Non-Maximum Suppression (NMS).
Experimental Results
SGCCNet was evaluated on the KITTI 3D Object Detection Benchmark.
- Accuracy: It achieved 80.82% AP3D for the "Moderate" Car category, surpassing IA-SSD by 2.35% and Fast Point R-CNN by 3.42%.
- Small Objects: Significant gains were observed for Pedestrians (+1%) and Cyclists (+3.1%) compared to previous state-of-the-art models.
- Efficiency: The model maintains an inference time of 23ms (~43 FPS), meeting the requirements for real-time autonomous systems.
- Generality: The SGDA and CCM modules are "plug-and-play," demonstrating performance boosts when integrated into other detectors like 3DSSD.
Conclusion
SGCCNet provides a robust solution for single-stage 3D object detection by addressing the fundamental issues of feature over-reliance and confidence misalignment. While the current implementation uses Farthest Point Sampling (FPS), which may face scaling challenges on extremely dense datasets, its performance on KITTI establishes it as a leading point-based detection framework.
Summary updated