Resource-Adaptivity Beyond the Model: Sensor Control for Quantized On-Device Vision

Resource-adaptive inference is typically framed as a model-side problem: reduce precision, prune computation, route inputs, or shrink architectures. For camera-based on-device vision, however, the sensor is also a controllable resource that determines the input evidence before inference. We study this interaction on datasets with environment and sensor shifts by evaluating a diverse set of vision models under full-precision and quantized inference across auto-exposure and adaptive sensor policies. Our results show that adaptive sensing often yields larger gains than increasing model size or spending more bits on it: in many cases, compressed models with adaptive sensing outperform higher-precision models using auto-exposure. These findings suggest that resource-adaptive inference for embodied vision should evaluate model--sensor pairs and treat sensing as part of the resource budget. We further show that a simple entropy-based model-free policy can recover much of this benefit without scoring every candidate image with the target model.