Reproduce “CF-YOLO: Cross Fusion YOLO for Object Detection in Adverse Weather with a High-quality Real Snow Dataset” on OpenResearch

About this paper

Qiqi Ding, Peng Li, Xuefeng Yan, Ding Shi, Luming Liang, Weiming Wang, Haoran Xie, Jonathan Li and 1 more

This research introduces RSOD, the first quantitatively evaluated and graded real-world snow object detection dataset, alongside CF-YOLO, a detector featuring a Cross Fusion block. CF-YOLO consistently outperforms state-of-the-art baselines on RSOD in snowy conditions while maintaining competitive performance on standard datasets like MSCOCO.

Key insights

• The Peak Act activation function and its unsupervised training strategy effectively isolate and quantify snow features, enabling objective grading of real-world snowy images by Snow Coverage Rate.
• Direct and simultaneous cross-scale feature fusion (as in the CF block) is more effective than sequential feature processing for mitigating information loss and enhancing object detection in snow-degraded images.
• High-quality real-world snowy datasets are crucial for training robust object detectors, as models trained solely on clean or synthetic data suffer significant performance degradation due to domain shift.

Summary by alphaXiv.