RTScale: Sensitivity-Aware Adaptive Image Scaling for Real-Time Object Detection
Open access
Date
2022Type
- Conference Paper
ETH Bibliography
yes
Altmetrics
Abstract
Real-time object detection is crucial in autonomous driving. To avoid catastrophic accidents, an autonomous car should detect objects with multiple cameras and make decisions within a certain time limit. Object detection systems can meet the real-time constraint by dynamically downsampling input images to proper scales according to their time budget. However, simply applying the same scale to all the images from multiple cameras can cause unnecessary accuracy loss because downsampling can incur a significant accuracy loss for some images. To reduce the accuracy loss while meeting the real-time constraint, this work proposes RTScale, a new adaptive real-time image scaling scheme that applies different scales to different images reflecting their sensitivities to the scaling and time budget. RTScale infers the sensitivities of multiple images from multiple cameras and determines an appropriate image scale for each image considering the real-time constraint. This work evaluates object detection accuracy and latency with RTScale for two driving datasets. The evaluation results show that RTScale can meet real-time constraints with minimal accuracy loss. Show more
Permanent link
https://doi.org/10.3929/ethz-b-000558309Publication status
publishedExternal links
Editor
Book title
34th Euromicro Conference on Real-Time SystemsJournal / series
Leibniz International Proceedings in Informatics (LIPIcs)Volume
Pages / Article No.
Publisher
Schloss Dagstuhl - Leibniz-Zentrum für InformatikEvent
Subject
Real-time object detection; Dynamic neural network execution; Adaptive image scaling; Autonomous driving; Self-driving carsMore
Show all metadata
ETH Bibliography
yes
Altmetrics