Introduction
In autonomous driving, LiDAR-based 3D object detection is essential for real-time environment perception. To manage the computational load of massive point clouds, researchers often employ "pillars"—vertical columns that collapse 3D points into a 2D grid. While efficient, this approach suffers from significant information loss and quantization challenges.
Standard Pillar Feature Encoders (PFEs) rely on max-pooling, which retains only the maximum feature value within a column. This effectively "flattens" the vertical structure, making it difficult to distinguish between objects with similar 2D footprints but different 3D geometries, such as a pedestrian versus a person standing on a vehicle. Furthermore, deploying these models on edge hardware typically requires INT8 quantization. Standard encoding schemes produce numerical distributions that vary widely, leading to substantial accuracy degradation during conversion.
Methodology
PillarHist is a novel, quantization-friendly feature encoder designed to replace traditional max-pooling with a height-aware histogram approach. This plugin module rethinks point summarization within a pillar to preserve vertical information while reducing computational overhead.
The PillarHist process involves:
- Height-aware Binning: Dividing each pillar's height into discrete bins.
- Histogram Generation: Counting the points within each height bin.
- Intensity Weighting: Incorporating LiDAR reflection intensity to create a weighted histogram, capturing semantic clues like material reflectivity.
- Pillar-Level Projection: Instead of running heavy Multi-Layer Perceptrons (MLPs) on every individual point, PillarHist performs a linear projection on the aggregated pillar-level histogram.
- Coordinate Integration: Concatenating global $(x, y)$ center coordinates to maintain spatial context.
Key Insights
- The Primacy of Height: Ablation studies indicate that the vertical ($z$) dimension is the most critical factor for 3D detection. Focusing on height distributions yields better results than complex spatial offsets in the $x/y$ plane.
- Max-Pooling Bottleneck: Visualizations reveal that max-pooling often focuses on the highest points (e.g., a car roof), ignoring the underlying geometric structure. PillarHist provides a more complete representation of object geometry.
- Quantization Stability: By utilizing histograms (counts and averages), the network input remains within a stable, predictable range, eliminating the "clipping errors" common in INT8 compression of 3D models.
Experimental Results
The researchers evaluated PillarHist on the nuScenes, KITTI, and Waymo Open Datasets.
- Accuracy: Integrating PillarHist into PointPillars and CenterPoint-Pillar yielded an average improvement of ~1.0 point in the nuScenes Detection Score (NDS).
- Efficiency: Moving MLPs to the pillar-level reduced inference latency by 4ms to 9ms on a V100 GPU. GFLOPs for the PFE stage dropped from 0.152 to 0.065 compared to the PointPillars baseline.
- Quantization Performance: Under INT8 quantization, standard PointPillars experienced a 7.3 NDS drop, whereas the PillarHist version lost only 1.6 NDS, demonstrating its suitability for edge deployment.
Conclusion
PillarHist offers a practical and efficient solution for enhancing pillar-based 3D detectors. While highly effective for pillars, future work may explore its applicability to voxel-based detectors and automated tuning of height bin parameters to suit varying LiDAR resolutions.
Summary updated