Nicola Storni

Semester Work
Supervisors: Dr. Radu Timofte, Christoph Mayer

Plant Species Mapping

In this work we test a method for weakly supervised image segmentation for the task of plant species localization and density estimation in drone images. We use a data set that consists of drone images along railways with image level labels that specify absence and presence of the plant species to train a classification network. Then, we use Class Activation Mapping (CAM) to extract heat maps for plant localization while training the classification model. We show that the heat maps are a rough segmentation estimate. The critical layer in the network is the feature aggregation layer that allows an image size independent classification network. Instead of always using global average pooling, we test and propose other feature aggregation layers. The aggregation layer based on global max pooling over a smoothed feature map achieves the best IoU of 67% on a pixel level annotated test set.