TG-fusion: A target-guided fusion framework for domain-adaptive counting

• Propose the use of feature fusion to unify contradictory information from different domains for domain adaptation. • Propose the TG-Fusion framework to integrate the target-domain features into the training in the source domain and infer on the target domain without the feature encoder of the source domain. • TG-Fusion is the first attempt to realise training time adaptation of ground-to-UAV and cross-type crop counting. • TG-Fusion demonstrates high robustness across different cross-domain counting adaptation tasks. Domain-adaptive counting aims to accurately estimate the distribution of the objects together with the counting numbers for an unlabelled target domain. In this paper, we propose using feature fusion to unify contradictory information from different domains for domain adaptation. This leads to a scene adaptation framework, Target-Guided Fusion (TG-Fusion), which simplifies and boosts the efficiency of cross-domain counting adaptation. The framework performs especially well with the proposed ground-to-UAV and cross-type crop counting adaptation, of which the datasets are for practical usage and have a small amount of data, yet distinctive inter-domain differences. The approach trains the network to consider the target scene when predicting for the source domain. This is realised by composing a flexible entangle functionality, which dynamically integrates the target features into the training in the source domain, while disentangling the target-domain stream for inference. Moreover, the ratio for the feature fusion is trainable throughout the training time and adjusted by the supervision of the source-domain labels and the similarity among different domains. Empirical experiments demonstrate that the proposed framework trains the model for better adaptation and shows high robustness across different cross-domain counting scenarios.