Source-linked AI summary
RangeDet:In Defense of Range View for LiDAR-based 3D Object Detection
Lue Fan, Xuan Xiong, Feng Wang, Naiyan Wang, Zhaoxiang Zhang
TL;DR
Pure range-view LiDAR detection has lagged behind other representations despite range view’s compactness, motivating RangeDet. The proposed framework addresses overlooked scale and geometric-coordinate issues with dedicated components, achieving comparable performance to state-of-the-art multi-view detectors and a 20 3D AP vehicle-detection margin over prior pure range-view methods.
Problem
Pure range-view object detectors substantially underperform voxelized, BEV, and multi-view methods despite range view’s compact representation and efficient processing.
Method
RangeDet is a pure range-view, single-stage anchor-free detector using Range Conditioned Pyramid, Meta-Kernel, weighted NMS, and range-view data augmentation.
Results
RangeDet achieves comparable performance with state-of-the-art multi-view detectors and surpasses prior pure range-view detectors by 20 3D AP in vehicle detection.
Takeaways & Limitations
Properly handling scale variation, geometric inconsistency, and proposal duplication can make pure range-view detection competitive with multi-view approaches.
Takeaways & Limitations
Range view is valid only for scans from one viewpoint and is unavailable for general point clouds with overlapping points.
Abstract
from arXiv · showhide
In this paper, we propose an anchor-free single-stage LiDAR-based 3D object detector -- RangeDet. The most notable difference with previous works is that our method is purely based on the range view representation. Compared with the commonly used voxelized or Bird's Eye View (BEV) representations, the range view representation is more compact and without quantization error. Although there are works adopting it for semantic segmentation, its performance in object detection is largely behind voxelized or BEV counterparts. We first analyze the existing range-view-based methods and find two issues overlooked by previous works: 1) the scale variation between nearby and far away objects; 2) the inconsistency between the 2D range image coordinates used in feature extraction and the 3D Cartesian coordinates used in output. Then we deliberately design three components to address these issues in our RangeDet. We test our RangeDet in the large-scale Waymo Open Dataset (WOD). Our best model achieves 72.9/75.9/65.8 3D AP on vehicle/pedestrian/cyclist. These results outperform other range-view-based methods by a large margin (~20 3D AP in vehicle detection), and are overall comparable with the state-of-the-art multi-view-based methods. Codes will be public.
1. Introduction
RangeDet addresses the underperformance of pure range-view LiDAR detection by targeting scale variation, 2D–3D coordinate inconsistency, and proposal duplication. Its compact range-view design achieves results comparable to multi-view detectors and improves vehicle detection over prior pure range-view methods.
- BEV introduces quantization error for distant objects, while point-view methods are difficult to scale efficiently to large autonomous-driving point clouds.
- Range view is compact, preserves the single-view point-cloud information, supports fast range-coordinate neighborhood queries, and can detect as far as the sensor permits.
- Existing pure range-view detectors remained more than 20 points 3D AP below state-of-the-art methods on WOD vehicle detection.
- Range-view detection must address scale variation between nearby and distant objects and inconsistency between 2D convolution coordinates and 3D outputs.
- RangeDet combines a Range Conditioned Pyramid, Meta-Kernel, weighted NMS, and transferred 3D data augmentation in a pure range-view, single-stage anchor-free framework.
- 20 3D AP is the reported vehicle-detection margin over previous pure range-view detectors, while the best model is comparable with state-of-the-art multi-view methods.
2. Related Work
Related work spans BEV, point-view, and range-view LiDAR detectors, with prior methods differing in representation, processing strategy, and proposal refinement.
- BEV-based 3D detectors: BEV detectors discretize 3D space into voxels or pillars, using learned or handcrafted features and sparse convolution for efficient processing.
- Point-view-based 3D detectors: Point-view detectors process unordered points using PointNet-based proposal generation, foreground segmentation, and box refinement, including both one- and two-stage designs.
- Range-view-based 3D detectors: Range-view detectors project point clouds into 2D range images and apply 2D convolutions to predict 3D bounding boxes.
3. Review of Range View Representation
A LiDAR range view organizes one scan as a 2D range image indexed by beam inclination and measurement azimuth. Its values encode geometric and return-signal information, but the representation is valid only from a single viewpoint.
- A LiDAR with m beams and n measurements forms an m × n range image for one scan cycle.
- Each range-image column shares an azimuth, each row shares an inclination, and pixel values contain range, intensity, and auxiliary information.
- The Cartesian point cloud is decoded from the range view’s spherical-coordinate representation.
- Range view is valid only for scans from one viewpoint because overlapping points may map to the same image pixel.
- WOD provides native range images with range, intensity, and elongation, where elongation measures laser-pulse-width extension useful for distinguishing spurious objects.
4. Methodology
RangeDet’s methodology combines range-conditioned feature pyramids, geometry-aware Meta-Kernel convolution, weighted NMS, range-view augmentation, and azimuth-aligned regression within a full-resolution range-view architecture.
- Range Conditioned Pyramid: Range Conditioned Pyramid assigns objects to feature layers according to range to address scale variation in range images.The pyramid follows FPN-like construction but changes object-to-layer assignment from image area to range.
- Meta-Kernel Convolution: Meta-Kernel generates location-adaptive convolution weights from relative Cartesian coordinates or other meta-information to encode local 3D structure.A shared MLP produces weight vectors for sampled neighbors, whose element-wise products are concatenated and aggregated by a 1 × 1 convolution or fully connected layer.
- Meta-Kernel Convolution: Meta-Kernel replaces channel-wise summation with concatenation and learned aggregation, exploiting the ordered n×n neighborhood of the range view.Unlike point-based operators with irregular neighborhoods and permutation-invariant aggregation, Meta-Kernel operates on regular 2D range-view neighborhoods.
- Weighted Non-Maximum Suppression: Weighted NMS averages high-overlap proposals using their confidence scores, enabled by dense proposal generation on the compact full-resolution range-view feature map.Proposals below 0.5 confidence are filtered, and proposals with IoU above 0.5 to the top-ranked box contribute to its weighted average.
- Data Augmentation in Range View: Range-view augmentation maps 3D rotations and flips to azimuth translations or vertical-axis flips while preserving point azimuth coordinates.Copy-Paste retains pasted objects’ original vertical pixel coordinates.
- Architecture: The overall architecture uses eight range-image channels, inserts Meta-Kernel in the second BasicBlock, and predicts through stride-1, stride-2, and stride-4 pyramid layers.Feature maps are downsampled to stride 16 and progressively upsampled to full resolution; regression targets use each point’s azimuth direction as local x-axis.
5. Experiments
Experiments on WOD evaluate RangeDet’s components, comparisons, long-range behavior, and runtime. Ablations support Meta-Kernel, Range Conditioned Pyramid, weighted NMS, and design choices for range-view detection.
- Study of Meta-Kernel Convolution: Meta-Kernel is more effective at lower-level stages, where feature layers have a closer association with geometric structure.
- Study of Meta-Kernel Convolution: Meta-Kernel significantly improves small-object detection by enhancing boundary information that can otherwise be diluted by background.
- Study of Meta-Kernel Convolution: Concatenation with a fully connected layer outperforms max-pooling or summation because range-view neighborhoods preserve relative ordering.
- Ablation Experiments: Meta-Kernel, Range Conditioned Pyramid, weighted NMS, and IoU prediction each improve RangeDet’s vehicle-detection performance in ablation experiments.
- RangeDet outperforms pure range-view methods and is slightly behind the state-of-the-art BEV-based two-stage method on WOD.
- Runtime Evaluation: 12 FPS is RangeDet’s runtime on a single 2080Ti GPU, without deliberate optimization, and its speed is unaffected by expanding valid detection distance.
6. Conclusion
RangeDet is a range-view-based detection framework built from Meta-Kernel, Range Conditioned Pyramid, and weighted NMS. With these designs, it achieves performance comparable to state-of-the-art multi-view detectors.
- RangeDet combines Meta-Kernel, Range Conditioned Pyramid, and weighted NMS in a range-view-based detection framework.
- RangeDet achieves comparable performance with state-of-the-art multi-view-based detectors.