Multi-Resolution CNN slashes latency, boosts safety in autonomous driving
A new CNN dynamically selects input resolution to balance speed and accuracy in real-time driving.
In autonomous driving, the tradeoff between latency and accuracy is critical for safety. A new paper from researchers Qitao Weng and Heechul Yun introduces a multi-resolution end-to-end CNN that addresses this challenge by allowing the model to dynamically choose an input resolution under a given latency budget. The system uses per-resolution batch normalization to handle multiple input scales, enabling smooth transitions between high-fidelity (high resolution, slower) and low-fidelity (low resolution, faster) inference without requiring access to the original training dataset. This flexibility is key for real-time cyber-physical systems where compute availability and scene complexity vary.
The model was evaluated in the CARLA urban driving simulator using monocular camera input. Compared to fixed-resolution baselines, the multi-resolution approach consistently improved safety metrics per route: fewer lane invasions, red-light violations, and collisions. The results demonstrate that adaptively trading resolution for latency can significantly enhance safety without sacrificing overall performance. This work, published as arXiv:2605.29138 and presented at ICCPS 2026, offers a practical pathway for deploying DNNs in latency-sensitive autonomous systems where worst-case timing matters as much as average accuracy.
- Uses per-resolution batch normalization to support multiple input scales from a single CNN without retraining.
- Achieves consistent reductions in lane invasions, red-light infractions, and collisions in CARLA simulations.
- Enables runtime resolution retargeting to optimize latency-accuracy tradeoff based on current compute and scene context.
Why It Matters
Real-world AVs need to adapt on-the-fly — this method cuts accidents by dynamically balancing speed and detail.