Multilabel Aerial Image Classification With Unsupervised Domain Adaptation

Multi Label Image Classification using Adaptive Graph Convolutional Networks (ML-AGCN)

“…However, similar to direct methods for traditional multi-label image classification, these two approaches [55], [57] neglect the important information of label dependencies. A graph-based approach called DA-MAIC has been then proposed as an alternative [22]. As in ML-GCN [10], they have proposed to build a graph for modeling label correlations based on label co-occurrences.…”

“…The second challenge is more general and concerns most multi-label classification techniques. They usually assume that unseen images and training data are drawn from the same distribution, hence ignoring a possible domain shift problem [22]. This leads, therefore, to poor generalization capabilities.…”

“…It aims at learning domain invariant features to bridge the gap between a source domain and a target domain without access to the associated labels. While unsupervised domain adaptation has been widely investigated in the context of single-label image classification, very few works have considered this strategy for multi-label image classification [22], [23]. One of the most well-known domain adaptation-based multi-label classification frameworks [22] leverages graph representations to model label inter-dependencies and couple it with an adversarial training approach [19].…”

“…While unsupervised domain adaptation has been widely investigated in the context of single-label image classification, very few works have considered this strategy for multi-label image classification [22], [23]. One of the most well-known domain adaptation-based multi-label classification frameworks [22] leverages graph representations to model label inter-dependencies and couple it with an adversarial training approach [19]. However, since it is a direct extension of [10], it naturally inherits from the same limitation mentioned earlier, e.g., (1) the heuristically defined structure of the graph, (2) an empirically set threshold, and (3) node feature dissimilarity.…”

Multi-label Aerial Image Classification Based on Image-Specific Concept Graphs

“…Recently, multi-label aerial image classification (MAIC) has attracted great attention and been widely used in practical projects, such as geomorphological monitoring [1], maritime traffic planning [2]. MAIC aims to predict multiple labels simultaneously in an image and explore fine-grained understandings [3,4]. Compared with traditional aerial image scene classification, MAIC is more challenging due to the higher image resolution and complicated object contents [5].…”