TL;DR
This paper introduces CP-loss, a novel loss function for road curb detection in aerial imagery that specifically addresses the common issue of disconnected curb segments in segmentation outputs. By leveraging skeletonization and a weighted combination of cross-entropy and Dice loss, CP-loss guides the network to produce more topologically correct and continuous road curb predictions, crucial for robust autonomous driving systems.
What Problem Does This Paper Address?
Autonomous driving systems rely heavily on accurately identifying road boundaries to ensure safe navigation. Road curbs, being thin, long, and often partially occluded, pose a significant challenge for traditional object detection methods. While existing approaches often use vehicle-mounted sensors, these are prone to severe occlusions in dynamic traffic. This paper proposes an offline approach using high-resolution aerial images, which are less susceptible to immediate traffic occlusions and can be used to build high-definition (HD) maps. The core problem addressed is the inherent disconnectivity issue in semantic segmentation outputs for line-shaped objects like road curbs. Standard loss functions like cross-entropy and Dice loss, while effective for general segmentation, fail to explicitly penalize broken curb predictions, leading to fragmented results that are difficult to use for downstream tasks.
Methods
The authors propose a novel loss function called CP-loss (Connectivity-preserving Loss). The core idea is to enhance semantic segmentation by explicitly considering the topological connectivity of road curbs. The process involves:
- Skeletonization: Both the predicted segmentation map (S) and the ground-truth label (GT) are skeletonized to obtain one-pixel-wide representations (Skelp and Skelg, respectively).
- Disconnection Identification: The difference between Skelp and Skelg is used to identify "failed-retrieved ground-truth skeleton pixels" (Skelfg) and "false-positive predicted skeleton pixels" (Skelfp). Skelfg highlights areas where the predicted curb is broken or missing, while Skelfp helps penalize spurious detections.
- Weighted Loss: CP-loss combines a weighted cross-entropy loss (LCE) and a weighted Dice loss (LDice). The weights (ui, vi, and βi) are dynamically calculated based on the Euclidean distance of pixels to Skelfg and Skelfp. Pixels closer to identified disconnections receive higher weights, forcing the network to focus on repairing these breaks. The distance-based weighting is combined with a Focal loss-like mechanism to prioritize harder-to-segment pixels.
- Semantic Segmentation Network: The CP-loss is designed to be plug-and-play and can be integrated with any standard semantic segmentation network. The authors use a UNet architecture for their experiments.
Datasets
The paper utilizes the NYC-planimetrics dataset provided by NYC OpenData. This dataset contains high-resolution aerial images with road curb annotations. The authors pre-process this dataset by:
- Splitting large images into smaller patches (1000x1000).
- Converting annotations from WGS84 to image coordinates.
- Filtering out patches with no curbs or inappropriate annotations. They created a custom dataset of approximately 18,000 image patches, with 17,000 used for training and 1,000 for testing.
Key Insights and Ideas
- Skeletonization as a Connectivity Indicator: The paper highlights that skeletonization is a powerful tool for capturing the structural topology of line-shaped objects, making it ideal for identifying and penalizing disconnectivity.
- Targeted Weighting for Disconnections: Instead of a uniform loss, CP-loss intelligently assigns higher weights to pixels in regions where disconnectivity is detected, effectively guiding the network's learning process towards producing more continuous outputs.
- Dual Supervision (Pixel & Image Level): Combining weighted cross-entropy (pixel-level) and weighted Dice loss (image-level) provides a more comprehensive supervision signal, capturing both local pixel accuracy and global structural consistency.
- Importance of Aerial Imagery for Offline Mapping: The use of high-resolution aerial images for offline curb detection offers a practical approach for generating HD maps, complementing real-time, vehicle-mounted sensor data.
Key Results
- Qualitative Improvement: Visualizations (Fig. 2 and Fig. 4) clearly show that CP-loss produces significantly more continuous and complete road curb skeletons compared to baseline loss functions.
- Quantitative Performance: In comparative experiments, CP-loss achieved the highest F1-score (0.8809) and SCM (0.7994), outperforming standard BCE, Balance CE, Distance CE, Focal loss, and Dice loss.
- Ablation Studies: The ablation studies confirmed the importance of both the weighted cross-entropy and weighted Dice loss components, as well as the learned weights themselves, for achieving optimal connectivity. Removing any of these significantly degraded performance.
Limitations and Future Work
- Failure Cases: The authors acknowledge that CP-loss cannot guarantee perfect connectivity in cases of severe occlusion (e.g., by trees or shadows). While it alleviates the issue compared to online methods, some failures persist.
- Model Architecture Independence: While CP-loss is designed to be plug-and-play, its effectiveness might be further enhanced by more advanced segmentation network architectures.
- Future Directions: The authors plan to:
- Further refine CP-loss to improve its effectiveness and efficiency.
- Design new network structures specifically incorporating CP-loss for road curb detection.
Summary updated