Efficient and Robust Object Detection against Multi-Type Corruption Using Complete Edge Based on Lightweight Parameter Isolation
- Authors
- Kim, Youngjun; Shin, Jitae
- Issue Date
- 1-Feb-2024
- Publisher
- Institute of Electrical and Electronics Engineers Inc.
- Keywords
- Autonomous-Driving; Catastrophic Forgetting; Feature extraction; Image edge detection; Meteorology; Multi-Type Corruption; Object detection; Parameter-Isolation; Robust Object Detection; Robustness; Semantics; Snow
- Citation
- IEEE Transactions on Intelligent Vehicles, v.9, no.2, pp 1 - 14
- Pages
- 14
- Indexed
- SCIE
SCOPUS
- Journal Title
- IEEE Transactions on Intelligent Vehicles
- Volume
- 9
- Number
- 2
- Start Page
- 1
- End Page
- 14
- URI
- https://scholarworks.bwise.kr/skku/handle/2021.sw.skku/110888
- DOI
- 10.1109/TIV.2024.3351271
- ISSN
- 2379-8858
- Abstract
- The objective of robust object detection against multi-type corruption (MTC) is to improve detection performance in various domains, including common risks in real-world applications, such as noise, blur, adverse weather, digital errors, and more. In driving scenarios, MTC can occur due to harsh conditions like tunnels, extended driving periods, high temperatures, object movements, and even desert driving. However, addressing MTC poses a significant challenge as it can lead to catastrophic forgetting (CF) issues due to domain shifts. To enhance robustness without encountering CF, we propose the unified stem block for corruption (USB-C). This approach overcomes CF by utilizing complete edge (CE), which not only represents the external edges of objects but also captures internal information. In addition, we introduce the DC-Converter, facilitating efficient CE extraction under MTC conditions using lightweight parameter isolation. Our method not only enhances robustness without CF-induced performance degradation but also achieves over 20 FPS on a single consumer-grade GPU with minimal additional memory usage. The performances are demonstrated through experiments on COCO-C and Bdd100k benchmarks. Moreover, our method demonstrates a model-agnostic property as it can be applied to two representative detectors, RetinaNet and Faster R-CNN. IEEE
- Files in This Item
- There are no files associated with this item.
- Appears in
Collections - Information and Communication Engineering > School of Electronic and Electrical Engineering > 1. Journal Articles
Items in ScholarWorks are protected by copyright, with all rights reserved, unless otherwise indicated.