Domain adapted broiler density map estimation using negative-patch data augmentation

To overcome the lack of labelled data in real-world situations such as agriculture, transfer learning has recently been in the spotlight. Domain adaptation is a transfer learning method that can migrate well learned knowledge from a source domain to a target domain. However, it is difficult to ensure the generalisation to adequate feature space when source and target domain are heterogeneous due to their excessive domain shift. In this study, a negative-patch (NP) data augmentation was proposed to improve the domain adaptation performance between heterogeneous domains for density map estimation. The proposed NP was applied to Domain-Adversarial Training of Neural Networks (DANN) for broiler flock density estimation. A people crowd dataset and a broiler dataset were used as source and target domain datasets, respectively. Experiments were conducted by the model configuration, and the broiler counting performances were compared. The results showed that the DANN with NP could reduce the error by up to 54% for DANN without NP and that the NP has demonstrated its generalized performance for multiple source domain datasets.