MeGA-CDA: Memory Guided Attention for Category-Aware Unsupervised Domain Adaptive Object Detection

Existing approaches for unsupervised domain adaptive object detection perform\nfeature alignment via adversarial training. While these methods achieve\nreasonable improvements in performance, they typically perform\ncategory-agnostic domain alignment, thereby resulting in negative transfer of\nfeatures. To overcome this issue, in this work, we attempt to incorporate\ncategory information into the domain adaptation process by proposing Memory\nGuided Attention for Category-Aware Domain Adaptation (MeGA-CDA). The proposed\nmethod consists of employing category-wise discriminators to ensure\ncategory-aware feature alignment for learning domain-invariant discriminative\nfeatures. However, since the category information is not available for the\ntarget samples, we propose to generate memory-guided category-specific\nattention maps which are then used to route the features appropriately to the\ncorresponding category discriminator. The proposed method is evaluated on\nseveral benchmark datasets and is shown to outperform existing approaches.\n