Object detection in hyperspectral images with unsupervised domain adaptation

Official URL: https://dx.doi.org/10.1109/UBMK63289.2024.10773548

Labeled data is essential for training deep learning models. When labeled data is scarce or unavailable, transferring knowledge from a model trained on a related task can help address this challenge and mitigate potential detrimental effects. However, when there is a distribution mismatch between the training data and the test data, this leads to decreased model performance, making it difficult to generalize. Domain adaptation techniques are one way to achieve robust and generalized infer-ences, especially when the target and source data distributions differ or when labeled target data is limited. This is particularly important in the context of hyperspectral data, where labeling is both challenging and costly, and labeled data is often extremely scarce, in contrast to the abundance of labeled data with optical images. In this study, an unsupervised domain adaptation approach is proposed for object detection, in order to adapt a model trained on optical source data to hyperspectral target data. The proposed method simultaneously employs adversarial alignment techniques at multiple levels of the model, effectively bridging the domain gap by aligning features across different scales of content description. The proposed approach is evaluated on the newly introduced M2S0DAI dataset, demonstrating promising results in improving object detection performance across domains.