BRASS: Budget-Aware RAW Sensor Sampling for Edge Vision via Co-Design

Kailash Talreja, Saurabh Jha

Keywords: RAW sensing, sensor–inference co-design, mixed-precision ADC, region-of-interest (ROI) readout, efficient vision, calibration (ECE), classification on RAW, edge computing, budget-aware learning, masked sampling, Imagenette-RAW, throughput measurement

TL;DR: BRASS learns where - and with how many ADC bits - to read RAW pixels, matching an RGB baseline at ~0.47× sensed-bit proxy with better calibration and higher throughput, illustrating sensor–model co-design for efficient edge vision.

Abstract: Most camera systems read every RAW pixel at uniform precision, wasting measurement budget on uninformative regions. We present BRASS, a budget-aware RAW sensing framework that treats sensed bits as a first-class resource. A tiny policy network makes per-patch decisions on (i) whether to read and (ii) the ADC bit-depth; a compact RAW backbone consumes the resulting sparse, mixed-precision tensor directly—without demosaicing. We train end-to-end with a budget-aware objective that exposes controllable accuracy–efficiency trade-offs. On Imagenette-RAW (160×160, RGGB), BRASS matches the accuracy of a small RGB baseline while using ≈0.47× the sensing-bit proxy and produces better-calibrated predictions (lower ECE, no temperature scaling). Synchronized A800 (batch=128) forward-pass timings show higher throughput, consistent with reduced sensed work (policy cost included; ISP/demosaic for RGB and host I/O excluded). Scope & limits: results are software-based; the sensing-bit proxy is not a full energy model; real deployment requires ROI readout and per-region ADC control, which are supported in modern CMOS sensors. BRASS illustrates Learning-to-Sense co-design by optimizing *what* to measure and *with how many bits* to measure it under explicit budgets. We will release training scripts, configs, and ONNX exports upon acceptance.

Submission Number: 8