Object detection (OD) plays a critical role in autonomous vehicle (AV) perception pipelines, but its performance often degrades under challenging environmental conditions such as adverse weather, varying lighting, and high scene complexity. These real-world variabilities introduce noise and contextual shifts that significantly impact both accuracy and system efficiency. Existing solutions primarily focus on architectural improvements or fine-tuning strategies for specific perturbations, often overlooking end-to-end pipelines that encompass data acquisition, model adaptation, and runtime inference optimization. In this work, we propose CoSense, a comprehensive framework for context and noise-aware object detection in AVs. Our system integrates efficient generative data collection to supplement hardto-acquire edge cases, targeted fine-tuning of detection models under noisy scenarios, and a lightweight heuristic for adaptive inference. By dynamically selecting models based on scene complexity and input quality, our framework achieves up to average 1.64x accuracy gain and 1.16x inference speedup compared to noise and context agnostic model selection baselines while keeping the end-to-end object detection latency under conventional tail latency threshold in autonomous driving applications.
Alikhani et al. (Tue,) studied this question.