Abstract
In this paper, we show that High-Definition (HD) maps provide strong priors that can boost the performance and robustness of modern 3D object detectors. Towards this goal, we design HDNET, a single-stage detector that extracts geometric and semantic features from HD maps. As maps might not be available everywhere, we also propose a map prediction module that estimates the map on the fly from raw LiDAR data. We conduct extensive experiments on KITTI as well as a large-scale 3D detection benchmark containing 1 million frames, and show that the proposed map-aware detector consistently outperforms the state-of-the-art in both mapped and un-mapped scenarios. Importantly, the entire framework runs at 20 frames per second.
Introduction
3D object detection is a fundamental task for autonomous driving, typically relying on LiDAR sensors to perceive the environment. However, LiDAR-based detection faces significant challenges due to point sparsity and occlusion, particularly at long ranges (50m+). Furthermore, variations in road geometry, such as slopes, introduce height offsets that complicate standard 3D convolutions. While HD maps are ubiquitous for motion planning, they are often underutilized in perception. HDNET addresses these limitations by treating HD maps as a "cheat sheet," incorporating geometric ground priors and semantic road masks directly into the detection pipeline to improve accuracy and robustness.
Methodology
HDNET is a fully convolutional single-stage detector that operates on a Bird’s Eye View (BEV) representation of LiDAR data. The core innovation lies in the fusion of map-derived priors with raw sensor data.
Geometric and Semantic Priors
To handle translation variance caused by road slopes, HDNET utilizes a Geometric Prior. Instead of absolute height ($z$), the model calculates height relative to the ground ($z - z_{ground}$), where $z_{ground}$ is retrieved from the HD map. This effectively "flattens" the world for the network. Additionally, a Semantic Prior is introduced by rasterizing the road layout from the HD map into a binary mask, which is fed as an extra input channel to guide the network toward regions where vehicles are likely to be located.
Online Map Estimation
Recognizing that HD maps are not always available, the authors developed an Online Map Estimation module. This U-Net-style sub-network estimates ground height and road masks on the fly from a single LiDAR sweep, allowing the detector to function in unmapped environments.
Key Insights
A critical advantage of HDNET is the "flattening" effect provided by the geometric prior. By normalizing height relative to the ground, the model does not need to learn object appearances at every possible elevation or slope, making features much more discriminative for sparse points at long ranges.
To prevent the model from becoming overly dependent on perfect map data, the researchers employed "map dropout" during training. By randomly omitting road masks, they ensured the detector remains reliable even when map data is noisy or unavailable.
Experimental Evaluation
The researchers evaluated HDNET on two major benchmarks: TOR4D, a massive internal dataset containing over 1 million frames, and KITTI.
- Accuracy Gains: On the TOR4D dataset, HDNET achieved a significant 5.49% AP gain at long ranges (50-70m) compared to the baseline.
- Generalization: When tested on KITTI using the online map estimator (trained on TOR4D), the model provided a 2.87% AP boost, demonstrating strong cross-continental generalization.
- Efficiency: The entire pipeline achieves a real-time speed of 20 FPS (50ms latency), making it suitable for deployment on autonomous vehicles.
Conclusion
HDNET demonstrates that integrating HD maps into the perception stack significantly enhances 3D object detection, particularly in challenging long-range scenarios. While online map estimation remains difficult beyond 40 meters due to sensor sparsity, the use of geometric and semantic priors provides a robust framework for safer and more accurate autonomous navigation.
Summary updated