Skip to main content

Unsupervised Domain Adaption with Adversarial Networks

Supervisor

Suitable for

MSc in Advanced Computer Science

Abstract

Co-supervised by Julian Godding - j.godding@gardin.co.uk At www.gardin.co.uk "Object counting and segmentation are necessary tasks in image-based plant phenotyping, notably for counting plant organs and segmenting different plant matter. One challenge for computer vision tasks in plant phenotyping is that, unlike large image datasets of general objects, plant image datasets usually include highly self-similar images with a small amount of variation among images within the dataset. An individual plant image dataset is often acquired under similar conditions (single crop type, same field) and therefore trying to directly use a CNN trained on a single dataset to new images from a different crop, field, or growing season will fail. This is because a model trained to count objects on one dataset (source dataset) will not perform well on a different dataset (target dataset) when these datasets have different prior distributions. This challenge is called domain-shift. An extreme case of domain shift in plant phenotyping is using a source dataset of plant images collected in an indoor controlled environment, which can be more easily annotated because plants are separated with controlled lighting on a blank background and attempting to apply the model to a target dataset of outdoor field images with multiple, overlapping plants with variable lighting and backgrounds, and blur due to wind motion. Giuffrida et al proposed a method to adapt a model trained to count leaves in a particular dataset to an unseen dataset in an unsupervised manner. Their method uses the representations of images from the source domain (from a pretrained network) to adapt to the target domain. They employed the Adversarial Discriminative Domain Adaptation approach to solving the domain shift problem in the leaf counting task across different datasets. Their method aligns the predicted leaf count distribution with the source domain’s prior distribution, which limits the adapted model to learning a leaf count distribution similar to the source domain. Ayalew et al propose a custom Domain-Adversarial Neural Network architecture and trained it using the unsupervised domain adaptation technique for density estimation-based organ counting. These approaches have delivered strong results in reducing the impact of domain shift. In this project, the data scientist will perform further research into the use of adversarial networks for domain adaptation and apply them to our unique plant image dataset. "