Abstract
Autonomous driving systems require highly accurate and robust semantic maps to navigate safely in diverse environments. However, standard models often fail when encountering out-of-distribution (OOD) conditions like extreme weather or sensor malfunctions. This report details Team Samsung-RAL's approach to the 2024 RoboDrive Challenge (Track 2), where they enhanced the BEVerse framework through backbone scaling, spatio-temporal fusion, and specialized data augmentations. Their methodology resulted in significant performance improvements on the RoboDrive benchmark, demonstrating increased resilience to real-world corruptions.
Introduction
Semantic map segmentation—identifying lanes, boundaries, and sidewalks from a Bird’s-Eye View (BEV)—is a cornerstone of modern autonomous navigation. While current state-of-the-art models perform well under ideal conditions, the 2024 RoboDrive Challenge highlights a critical vulnerability: robustness to "natural corruptions." These include environmental factors like fog, snow, and rain, as well as hardware issues like motion blur and sensor noise. Team Samsung-RAL addressed these challenges by focusing on three key areas: feature extraction strength, temporal consistency, and training-time data diversity.
Methodology
The proposed solution is built upon the BEVerse multi-task framework. The team introduced three primary enhancements to improve OOD robustness:
1. Stronger Backbones
The team replaced standard image encoders with Swin Transformers (Tiny, Small, and Base). Larger backbones are hypothesized to learn more generalized and noise-resistant features compared to smaller, more specialized encoders.
2. Spatio-Temporal Fusion
To mitigate the impact of transient corruptions (e.g., a camera being momentarily obscured), the model incorporates temporal information from $N$ previous frames. This allows the system to maintain a consistent map by "remembering" features from clear timestamps.
3. Robustness-Oriented Data Augmentation
A suite of augmentations was integrated into the training pipeline to "vaccinate" the model against expected test-time corruptions. These include GridMask for simulating occlusions, Color Jittering for weather/lighting variations, and Channel Swapping to handle sensor color shifts.
Results and Discussion
The effectiveness of these strategies was evaluated using the nuScenes dataset for training and the RoboDrive Benchmark for testing.
Backbone Scaling and Temporal Fusion
Scaling the backbone from Swin-Tiny to Swin-Base provided a massive jump in performance, increasing mean Intersection-over-Union (mIoU) by 7.77 points on the Phase 1 test set. Adding the spatio-temporal fusion module further boosted the Base model's performance from 21.68 to 24.08 mIoU, confirming that temporal context acts as a vital safety net.
Impact of Data Augmentation
The augmentation strategy proved particularly effective for smaller models. On the Phase 2 test set, applying the full augmentation suite to the Tiny model increased its mIoU from 23.24 to 29.76. This suggests that targeted training can significantly narrow the robustness gap between small and large models.
Conclusion
Team Samsung-RAL's technical strategy underscores the importance of model scale, temporal context, and diverse training data in achieving OOD robustness. Future work will focus on combining these elements—applying the full augmentation suite to the largest backbones—and exploring automated augmentation techniques to further harden perception systems for autonomous driving.
Summary updated