Multi-Resolution End-to-End Deep Neural Network for Optimizing Latency-Accuracy Tradeoff in Autonomous Driving

Researchers present a multi-resolution deep neural network for autonomous driving that dynamically selects input resolution based on latency constraints and compute availability. The approach uses per-resolution batch normalization and resolution retargeting to optimize the tradeoff between prediction accuracy and processing speed, demonstrating improved safety metrics in CARLA simulations compared to fixed-resolution models.

This research addresses a critical engineering challenge in autonomous driving: the fundamental tension between model accuracy and real-time performance. Traditional fixed-resolution neural networks operate at a single point on the latency-accuracy spectrum, proving suboptimal when environmental conditions or computational resources fluctuate. The proposed multi-resolution architecture solves this by enabling dynamic adaptation, allowing the same model to operate efficiently across varying hardware constraints and driving scenarios.

The innovation centers on technical elegance rather than raw capability gains. Per-resolution batch normalization permits a single network to maintain performance across multiple input scales, while resolution retargeting enables training on diverse image resolutions without requiring access to original datasets—a practical advantage for researchers and practitioners facing data access constraints. This approach reflects broader industry trends toward efficient, adaptive AI systems that maximize utility within computational budgets.

For autonomous vehicle developers, this architecture reduces deployment friction by eliminating the need to maintain separate models for different hardware platforms or environmental conditions. Safety-critical applications benefit from the consistent improvements in lane-keeping, traffic law compliance, and collision avoidance demonstrated across CARLA's urban scenarios. The work validates that computational efficiency and safety need not be opposing forces.

Future development should focus on real-world validation beyond simulation, particularly testing performance degradation under sensor noise and adverse weather. Integration with production autonomous systems requires proving the adaptive mechanism responds appropriately during edge cases where latency budgets compress unexpectedly. Standardization of multi-resolution evaluation metrics across the industry would accelerate adoption of similar adaptive approaches.

• →Multi-resolution networks dynamically optimize latency-accuracy tradeoffs by selecting appropriate input resolution during inference based on compute constraints.
• →Per-resolution batch normalization enables a single CNN to maintain performance across multiple image scales without separate model variants.
• →Resolution retargeting permits multi-resolution training without access to original datasets, reducing data engineering overhead.
• →CARLA evaluation shows consistent safety metric improvements in lane keeping, traffic compliance, and collision avoidance versus fixed-resolution baselines.
• →The approach addresses a critical challenge for autonomous vehicle deployment across heterogeneous hardware platforms and real-world conditions.