FG-DINO: Inducing Foreground Approximation Into DINO for Low-Resolution Surveillance Videos

The predominantly constant image values in surveillance videos provide a very stable image background, enabling a straightforward extraction of moving objects such as persons or cars. This unique property is not extensively exploited by modern video object detectors due to their focus on video datasets with moving cameras. We rethink the task of object detection in surveillance videos based on the integration of simple temporal features into modern object detectors. The integration of a foreground approximation supports the object detector with easily extractable temporal information. The foreground information does not risk significant overfitting due to the high input dimension of multiple frames present in most video object detectors. In contrast, our FG-DINO approach only concatenates one input channel and still shows significant advantages in terms of generalization performance, even in cross-dataset evaluations. Our code contains the necessary modifications: https://github.com/FellowPlanter/FG-DINO.