RedunCut: Measurement-Driven Sampling and Accuracy Performance Modeling for Low-Cost Live Video Analytics

TL;DR

RedunCut tackles the high cost of live video analytics by introducing a measurement-driven sampler and a data-driven accuracy predictor for dynamic model-size selection. It explicitly estimates the cost-benefit of sampling and uses a lightweight kNN-based predictor to map prediction statistics to accuracy, all without modifying base models. Across road-vehicle, drone, and surveillance videos and multiple model families, RedunCut achieves 14-62% compute savings at the same accuracy, and shows robustness to limited historical data and data drift. This work demonstrates a practical, model-agnostic shim that enables cost-efficient LVA at scale.

Abstract

Live video analytics (LVA) runs continuously across massive camera fleets, but inference cost with modern vision models remains high. To address this, dynamic model size selection (DMSS) is an attractive approach: it is content-aware but treats models as black boxes, and could potentially reduce cost by up to 10x without model retraining or modification. Without ground truth labels at runtime, we observe that DMSS methods use two stages per segment: (i) sampling a few models to calculate prediction statistics (e.g., confidences), then (ii) selection of the model size from those statistics. Prior systems fail to generalize to diverse workloads, particularly to mobile videos and lower accuracy targets. We identify that the failure modes stem from inefficient sampling whose cost exceeds its benefit, and inaccurate per-segment accuracy prediction. In this work, we present RedunCut, a new DMSS system that addresses both: It uses a measurement-driven planner that estimates the cost-benefit tradeoff of sampling, and a lightweight, data-driven performance model to improve accuracy prediction. Across road-vehicle, drone, and surveillance videos and multiple model families and tasks, RedunCut reduces compute cost by 14-62% at fixed accuracy and remains robust to limited historical data and to drift.

Figures (23)

• Figure 1: The accuracy of various model sizes in a 20 second video (top), and the resulting model size selection (bottom). Top: The accuracy (y-axis) of different model sizes (color) on 2-second video segments of a road vehicle video (x-axis). Bottom: The resulting model size selection (y-axis) for each video segment (x-axis). The models are taken from the EfficientDet (EDet) efficientdet family.
• Figure 2: Periodic sampling and selection.
• Figure 3: The cost of various model size selection methods in different video workload types and accuracy targets. The cost (y-axis) of various model size selection methods (color, bar pattern), normalized to the cost of the sample-once method (red) in surveillance (left panel) and road vehicle camera (right panel) video workloads across several accuracy targets (x-axis). Bars with patterns (e.g. Tiered-sampling (high-sample)), represent hypothetical solutions.
• Figure 4: Example for the use of historical workload data to estimate components (i) and (ii). Top left: The costs of sampling and selection. Top center: Predicted accuracy when no models are sampled. Bottom: Varying outcomes of sampling Large, based on the prediction statistic value in the video segment. See text for details.
• Figure 5: Using the historical workload data (filtered and unfiltered) to get the range of likely values of prediction statistics. The histogram of the prediction statistic (sum of objectness score ddsEAAR) of model EDet-D7X efficientdet in different video segments of the historical workload data (road vehicles videos waymo). The histogram is shown unfiltered for all video segments (blue), and filtered: only for segments where EDet-D5=3 (orange), only for segments where EDet-D3=12 (red), and only for segments where EDet-D5=3 and also EDet-D7=4 (green).