SCIENCE@home

Data Augmentation is All You Need for Robust Fisheye Object Detection

L. Pham,Quoc Pham-Nam Ho,Duong Khac Vu,Huy-Hung Nguyen,Chi Dai Tran,Duong Nguyen-Ngoc Tran,T. H. Tran,Ngoc Doan-Minh Huynh,Hyung-Joon Jeon,Hyung-Min Jeon,Son Hong Phan,Trinh Le Ba Khanh,Jae Wook Jeon

Published 2025 in 2025 IEEE/CVF International Conference on Computer Vision Workshops (ICCVW)

ABSTRACT

Fisheye cameras are vital for traffic surveillance systems (TSSs) due to their wide field-of-view, but image distortions challenge accurate object detection. Conventional models like YOLOs and RT-DETR struggle with fisheye images due to limited datasets. This paper proposes several data augmentation techniques to enhance data efficiency by generating high-quality annotations from open-source TSS datasets. Particularly, we propose a set of augmentation techniques to simulate fisheye images, balance class distributions, and bridge the gap between day-night scenes. We also develop a real-time object detection framework using the state-of-the-art DEIM model, optimized with mixed precision quantization for deployment on edge devices. Our framework achieves strong performance in the AI City Challenge 2025, securing the 5th place on the F1-score ranking. Real-world experiments on an NVIDIA Jetson AGX Orin 32GB confirm real-time performance exceeding 25 FPS with minimal accuracy reduction. Extensive experiments using the harmonic mean of F1-score and FPS also yield satisfactory results. The code and data are available at GitHub11github.com/SKKUAutoLab/AIC25_Track_04.

PUBLICATION RECORD

CITATION MAP

EXTRACTION MAP

CLAIMS

  • No claims are published for this paper.

CONCEPTS

  • No concepts are published for this paper.

REFERENCES

Showing 1-32 of 32 references · Page 1 of 1

CITED BY