ROMA: Run-Time Object Detection To Maximize Real-Time Accuracy

TL;DR

ROMA addresses the challenge of maintaining high real-time detection accuracy when video content and compute resources vary. It introduces a run-time accuracy variation model that estimates Relative Average Precision (RAP) between detectors without ground-truth labels, leveraging runtime cues such as object size histograms and detection latency. The method combines offline AP estimation, an AP degradation model across dropped frames via a degradation factor $\beta$, and a RAP-based detector selection mechanism to switch among detectors in real time. Experiments on MOT17Det and MOT20Det with four YOLOv4 variants on an NVIDIA Jetson Nano demonstrate substantial real-time accuracy gains over single detectors and prior runtime techniques, highlighting ROMA's practicality for dynamic, resource-constrained video analytics.

Abstract

This paper analyzes the effects of dynamically varying video contents and detection latency on the real-time detection accuracy of a detector and proposes a new run-time accuracy variation model, ROMA, based on the findings from the analysis. ROMA is designed to select an optimal detector out of a set of detectors in real time without label information to maximize real-time object detection accuracy. ROMA utilizing four YOLOv4 detectors on an NVIDIA Jetson Nano shows real-time accuracy improvements by 4 to 37% for a scenario of dynamically varying video contents and detection latency consisting of MOT17Det and MOT20Det datasets, compared to individual YOLOv4 detectors and two state-of-the-art runtime techniques.

Figures (7)

• Figure 1: Accuracy Variation with Dynamically Varying Objects' Speeds and Available Compute Resources
• Figure 2: Notations for Frames and Frame Block Sizes
• Figure 3: Average APs (MOT17Det and MOT20Det)
• Figure 4: MOT17-04 (Left) and MOT20-05 (Right) MOTDet
• Figure 5: Average APs across All Cases (MOT17Det+MOT20Det)