Leaf Counting Without Annotations Using Adversarial Unsupervised Domain Adaptation

Abstract

Deep learning is making strides in plant phenotyping and agriculture. But pretrained models require significant adaptation to work on new target datasets originating from a different experiment even on the same species. The current solution is to retrain the model on the new target data implying the need for annotated and labelled images. This paper addresses the problem of adapting a previously trained model on new target but unlabelled images. Our method falls in the broad machine learning problem of domain adaptation, where our aim is to reduce the difference between the source and target dataset (domains). Most classical approaches necessitate that both source and target data are simultaneously available to solve the problem. In agriculture it is possible that source data cannot be shared. Hence, we propose to update the model without necessarily sharing the data of the training source to preserve confidentiality. Our major contribution is a model that reduces the domain shift using an unsupervised adversarial adaptation mechanism on statistics of the training (source) data. In addition, we propose a multi-output training process that (i) allows (quasi-)integer leaf counting predictions; and (ii) improves the accuracy on the target domain, by minimising the distance between the counting distributions on the source and target domain. In our experiments we used a reduced version of the CVPPP dataset as source domain. We performed two sets of experiments, showing domain adaptation in the intra- and inter-species case. Using an Arabidopsis dataset as target domain, the prediction results exhibit a mean squared error (MSE) of 2.3. When a different plant species was used (Komatsuna), the MSE was 1.8.