Cross-domain few-shot video object detection based on data augmentation
Aiming at the problems of scarce samples and large domain differences in cross-domain few-shot video object detection under non-cooperative scenarios, we propose a cross-domain few-shot video object detection method based on data augmentation. Taking advantage of the characteristics of small background changes and a small number of targets in the video to be detected, the method inherits the “copy-paste” strategy and combines the Poisson fusion algorithm to bidirectionally transplant and fuse the targets in the relevant annotated images with those in the dataset images, generating domain-consistent synthetic samples. This not only expands the amount of training data but also achieves the unification of target features between the source domain and the target domain. Experimental results show that the proposed method effectively alleviates the problems of sample scarcity and domain shift, is simple and efficient, and is compatible with various target detection models. It can provide a feasible technical path for cross-domain few-shot video object detection under non-cooperative scenarios.