DuEqNet: Dual-Equivariance Network in Outdoor 3D Object Detection for Autonomous Driving

Xihao Wang, JiaMing Lei, Hai Lan, Arafat Al-Jawari, Xian Wei

Abstract

Outdoor 3D object detection is a cornerstone of environment perception for autonomous driving. In complex traffic scenarios, precise object recognition is vital for the prediction and planning modules of dynamic systems, directly impacting safety and reliability. However, as vehicles veer and turn, the constant rotation of the surrounding environment relative to the sensors poses a significant challenge for perception systems. Most existing methods do not explicitly address the detection accuracy impairment caused by vehicle rotation. In this paper, we introduce DuEqNet, the first framework to incorporate the concept of equivariance into 3D object detection via a hierarchical embedded structure. DuEqNet extracts equivariant features at both local and global levels. Locally, a graph-based strategy ensures feature equivariance within point cloud pillars. Globally, group equivariant convolution layers aggregate these local features to achieve global equivariance. Evaluated on the nuScenes benchmark, DuEqNet achieves state-of-the-art performance, demonstrating superior orientation accuracy and prediction efficiency. Furthermore, our dual-equivariance strategy exhibits a "plug-and-play" capability, enhancing the performance of various popular object detection frameworks.

Introduction: The Rotation Challenge

In the realm of autonomous driving, the "ego-vehicle" is in a state of constant motion, frequently changing its heading. This movement causes the orientation of the surrounding environment to shift continuously relative to the vehicle's LiDAR sensors. Standard deep learning architectures for 3D object detection often struggle to maintain consistency under these conditions because they lack rotation equivariance.

Equivariance is a mathematical property where a transformation of the input (such as a rotation) results in an equivalent transformation of the network's internal representations. Without this property, a model might accurately detect a vehicle facing north but fail to recognize the same vehicle if it is rotated 45 degrees. While many developers rely on heavy data augmentation—training models on millions of rotated samples—this is a computationally expensive "brute-force" solution that does not address the underlying architectural inability to handle geometric transformations.

Methodology: The Dual-Equivariance Framework

DuEqNet addresses this limitation by embedding rotation handling directly into the network architecture through a two-tiered hierarchical approach.

1. Local Equivariance (Pillar-level)

Modern 3D detectors typically discretize LiDAR point clouds into vertical columns called "pillars." DuEqNet treats the points within each pillar as a graph. Drawing inspiration from directional message-passing networks used in molecular chemistry, the model employs a specialized embedding strategy. By utilizing filters based on the relative distances and positions of neighboring points, the network ensures that the features extracted from a single pillar remain invariant to rotation. This provides a stable foundation for the higher-level spatial features.

2. Global Equivariance (BEV-level)

Once the pillars are processed into a 2D pseudo-image, known as the Bird's Eye View (BEV), the model applies Group Equivariant Convolutions (G-CNNs). Specifically, DuEqNet utilizes the $P_4$ symmetry group, which accounts for translations and 90-degree rotations. This process involves:

  • A Lifting Layer: This expands the standard feature map to include rotation dimensions.
  • Equivariant Convolutions: These layers maintain symmetry throughout the backbone, ensuring that the spatial arrangement of pillars is understood regardless of the vehicle's orientation.

The network concludes with a center-based detection head (similar to CenterPoint) to regress final bounding boxes, orientations, and velocities.

Experimental Results

We evaluated DuEqNet on the nuScenes dataset, a large-scale benchmark featuring 40,000 key frames and 10 object classes. The model was compared against several high-performing baselines using metrics such as mean Average Precision (mAP), nuScenes Detection Score (NDS), and mean Average Orientation Error (mAOE).

Performance Highlights

  • State-of-the-Art Accuracy: DuEqNet achieved a mAP of 50.49% and an NDS of 60.6% on the nuScenes validation set, outperforming established models like CenterPoint and PointPainting.
  • Orientation Precision: The most significant gain was observed in orientation prediction. The model achieved an mAOE of 0.3506, representing an 8.9% relative improvement over previous best-performing methods. This confirms that global equivariance is crucial for understanding the heading of objects that span multiple pillars.
  • Efficiency: By making the architecture mathematically aware of rotation, DuEqNet achieves these results without the massive computational overhead required by exhaustive rotation-based data augmentation.

Key Insights and Generalization

A primary finding of this research is the "plug-and-play" nature of the DuEqNet modules. The authors integrated the dual-equivariance structure into other popular frameworks, including PointPillars and CenterPoint. In every instance, the addition of equivariant modules led to a measurable increase in detection performance and orientation accuracy.

Qualitative analysis further shows that DuEqNet significantly reduces "leak" detections (false positives in empty space) and improves the stability of bounding boxes during ego-vehicle maneuvers compared to non-equivariant models.

Conclusion and Future Work

DuEqNet demonstrates that hierarchical equivariance is a powerful tool for 3D perception in dynamic environments. By addressing rotation at both the local pillar level and the global BEV level, the model achieves superior orientation accuracy and robust detection.

Future research will look toward expanding global equivariance from discrete 90-degree increments ($P_4$ group) to continuous rotation equivariance to handle arbitrary angles with even greater precision. Additionally, the team aims to extend this dual-equivariance concept to multi-modal sensor fusion, combining the strengths of LiDAR and camera data in a unified, equivariant framework.

Summary updated

Abstract

Outdoor 3D object detection has played an essential role in the environment perception of autonomous driving. In complicated traffic situations, precise object recognition provides indispensable information for prediction and planning in the dynamic system, improving self-driving safety and reliability. However, with the vehicle's veering, the constant rotation of the surrounding scenario makes a challenge for the perception systems. Yet most existing methods have not focused on alleviating the detection accuracy impairment brought by the vehicle's rotation, especially in outdoor 3D detection. In this paper, we propose DuEqNet, which first introduces the concept of equivariance into 3D object detection network by leveraging a hierarchical embedded framework. The dual-equivariance of our model can extract the equivariant features at both local and global levels, respectively. For the local feature, we utilize the graph-based strategy to guarantee the equivariance of the feature in point cloud pillars. In terms of the global feature, the group equivariant convolution layers are adopted to aggregate the local feature to achieve the global equivariance. In the experiment part, we evaluate our approach with different baselines in 3D object detection tasks and obtain State-Of-The-Art performance. According to the results, our model presents higher accuracy on orientation and better prediction efficiency. Moreover, our dual-equivariance strategy exhibits the satisfied plug-and-play ability on various popular object detection frameworks to improve their performance.
BibTeX
@inproceedings{wang2023dueqnet,
  author = {Xihao Wang and JiaMing Lei and Hai Lan and Arafat Al-Jawari and Xian Wei},
  title = {{DuEqNet: Dual-Equivariance Network in Outdoor 3D Object Detection for Autonomous Driving}},
  booktitle = {IEEE International Conference on Robotics and Automation},
  year = {2023},
  url = {https://openreview.net/forum?id=DYxo4SM6ZF},
}
Figure