Abstract
Accurate 3D object detection in point clouds is a cornerstone of autonomous driving, yet it remains hampered by the "label assignment" problem—the process of determining which spatial locations represent positive samples during training. Traditional anchor-based and center-based methods often suffer from sample imbalance between large and small objects or inefficient use of high-quality data points. We introduce DCDet, a novel framework that addresses these challenges through two key innovations: Dynamic Cross Label Assignment (DCLA) and Rotation-Weighted IoU (RWIoU). DCLA employs a cross-shaped sampling region to adaptively select positive samples, ensuring balanced learning across object scales. RWIoU simplifies the computationally expensive 3D box overlap calculation by decoupling rotation into a weighting factor. Evaluated on the Waymo Open Dataset and KITTI, DCDet demonstrates state-of-the-art performance, particularly for small objects like pedestrians and cyclists, even with limited training data.
Introduction
In 3D object detection, models must categorize and locate objects within sparse point clouds. A critical training step is label assignment: deciding which grid cells or anchors should be treated as "positive." Existing methods generally fall into two categories: anchor-based methods, which create an imbalance by assigning far more anchors to large objects (e.g., trucks) than small ones (e.g., pedestrians); and center-based methods, which focus only on the object's center, potentially ignoring valuable features nearby. Furthermore, calculating the Intersection over Union (IoU) for rotated 3D boxes is mathematically complex and often leads to unstable optimization. DCDet is designed to overcome these inefficiencies.
Methods
Dynamic Cross Label Assignment (DCLA)
To bridge the gap between anchor-based and center-based approaches, we propose DCLA. Instead of a single center point or a rigid anchor grid, DCLA utilizes a cross-shaped sampling region centered on the object. This shape captures the most informative features along an object's principal axes without the redundancy of a full square or cubic region.
As shown in the pipeline above, DCLA operates through a "Sort & Select" mechanism. The model calculates a selection cost—a combination of classification and regression error—for points within the cross. It then adaptively determines the number of positive samples ($k$) based on prediction quality. This dynamic approach ensures that small objects receive sufficient positive samples for effective learning while preventing large objects from dominating the training signal.
Rotation-Weighted IoU (RWIoU)
Standard 3D IoU calculations are computationally heavy due to the geometry of intersecting rotated boxes. DCDet introduces RWIoU to decouple rotation from the spatial overlap calculation.
The RWIoU process involves four steps:
- Aligning axes of the predicted and ground-truth boxes.
- Calculating the intersecting volume ($V_{inter}$) as if the boxes were axis-aligned.
- Applying a rotation weighting factor ($\omega$) derived from the angular difference ($\Delta\theta$).
- Computing the final RWIoU score.
This method provides a smooth, differentiable loss function that effectively penalizes rotation errors without the instability of traditional 3D geometric intersections.
Experimental Results
DCDet was evaluated on the Waymo Open Dataset and KITTI benchmarks. On the Waymo validation set, DCDet achieved 75.0 mAP (Level 2), surpassing previous state-of-the-art methods like DSVT (73.2 mAP).
Notably, the model showed significant mastery over small objects. On the Waymo test set, DCDet delivered substantial gains in the Pedestrian (+5.1 mAP) and Cyclist (+5.5 mAP) categories compared to the PV-RCNN++ baseline. Furthermore, the framework demonstrated exceptional data efficiency; DCDet trained on only 20% of the Waymo dataset outperformed several competitive models trained on the full 100% of the data.
Conclusion
DCDet demonstrates that the strategy for label assignment is as critical as the volume of training data. By utilizing a dynamic cross-shaped sampling region and a simplified rotation-weighted loss, DCDet achieves superior accuracy and efficiency. Future work may explore deformable sampling regions to better adapt to complex object geometries and extreme occlusions.
Summary updated