Feature Extraction using Downsampling for Person Re-identification with Low-Resolution Images

We downsample higher-resolution pedestrian images to match lower-resolution images and extract features for person re-identification. This makes use of low-frequency image information without requiring paired high- and low-resolution training images.

Person re-identification means checking whether people shown by different cameras are the same individual. A person farther from a camera appears in less detail, making the image difficult to compare with a sharper image taken nearby. Many learning methods need numerous pairs of sharp and low-resolution images of the same people.

We consider situations where these paired training images cannot be collected. Our approach reduces the resolution of sharper images to match the coarser ones, a process called downsampling. It then extracts information for matching people. Rather than trying to recreate missing details, the approach uses the broad changes in color and brightness that remain in both images.

We collected images while changing the lens focal length and the distance between the person and camera. In our experiments, reducing sharper images produced more accurate person matching than enlarging low-resolution images. The results also confirmed that broad image patterns, not only fine details, contain useful information for distinguishing people.

• Masashi Nishiyama, Takuya Endo, Yoshio Iwai,
  Feature Extraction using Downsampling for Person Re-identification with Low-Resolution Images, Proceedings of 17th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP), Vol. 5: VISAPP, pp. 351 - 358, February 2022.
• 遠藤 拓弥, 西山 正志, 岩井 儀雄,
  低解像度に劣化した画像を用いた人物対応付けにおけるダウンサンプリング処理の効果検証,
  電子情報通信学会論文誌 A, Vol. J103-A, No. 12, pp. 289 - 298, December 2020.
• Keiji Obara, Hiroki Yoshimura, Masashi Nishiyama, Yoshio Iwai,
  Low-resolution Person Recognition using Image Downsampling, Proceedings of 15th IAPR International Conference on Machine Vision Applications (MVA), pp. 448 - 451, May 2017.