Methods and metrics for efficient video object detection
Abstract/Contents
- Abstract
- Understanding video contents by object detection is an essential task for machine intelligence. Video is an abundant data source, yet detecting objects from videos is computationally intensive or even prohibitive on edge devices. We start with efficient video object detection methods, which require fewer average FLOPs per video frame. We also find out that current widely adopted metrics do not fully describe a video object detector; therefore, we propose new metrics that measure the swiftness of an algorithm. By combining the proposed methods and evaluating them with the new metrics, we design a video object detection system that achieves high average accuracy, low response time, and a balanced workload under a limited computation budget. Two efficient methods for video object detection described in this thesis are CaTDet, which reduces the spatial area of detection, and PatchNet, which reduces the detection frequency with little accuracy drop. Experiments on multiple datasets show that CaTDet and PatchNet can reduce computation by 3.8x-13.0x and 3.4-4.9x, respectively. Combining CaTDet and PatchNet, we design SwiftDet, which performs joint spatial-temporal reduction and reduces FLOPs by 6.0x on ImageNet for Faster R-CNN ResNet-101. Evaluation with average delay and average latency further demonstrates that SwiftDet achieves a lower response time than CaTDet or PatchNet alone.
Description
Type of resource | text |
---|---|
Form | electronic resource; remote; computer; online resource |
Extent | 1 online resource. |
Place | California |
Place | [Stanford, California] |
Publisher | [Stanford University] |
Copyright date | 2021; ©2021 |
Publication date | 2021; 2021 |
Issuance | monographic |
Language | English |
Creators/Contributors
Author | Mao, Huizi |
---|---|
Degree supervisor | Dally, William J |
Degree supervisor | Horowitz, Mark (Mark Alan) |
Thesis advisor | Dally, William J |
Thesis advisor | Horowitz, Mark (Mark Alan) |
Thesis advisor | Fatahalian, Kayvon |
Degree committee member | Fatahalian, Kayvon |
Associated with | Stanford University, Department of Electrical Engineering |
Subjects
Genre | Theses |
---|---|
Genre | Text |
Bibliographic information
Statement of responsibility | Huizi Mao. |
---|---|
Note | Submitted to the Department of Electrical Engineering. |
Thesis | Thesis Ph.D. Stanford University 2021. |
Location | https://purl.stanford.edu/qw784gm7998 |
Access conditions
- Copyright
- © 2021 by Huizi Mao
Also listed in
Loading usage metrics...