Research
My research is organized around the directions below.
Interpretable visual recognition
Models that explain their predictions in terms people can check, such as object parts and visual concepts. A main use case is fine-grained recognition, where experts rely on specific traits to tell similar species apart.
Key publications
PDiscoFormer: relaxing part discovery constraints with vision transformers
Coarse-to-fine concept bottleneck models
Biodiversity monitoring
Machine learning for species identification and species distribution modelling, from citizen-science photos to field data, to support mapping and monitoring of biodiversity.
Key publications
Applying the maximum entropy principle to neural networks enhances multi-species distribution models
Effective integration of drone technology for mapping and managing palm species in the Peruvian Amazon
Multimodal Earth observation
General models for satellite and aerial imagery that combine several sensors and data types, including text, and transfer to new tasks in agriculture and environmental monitoring.
Key publications
Atomizer: generalizing to new modalities by breaking satellite images down to a set of scalars
Learning deep structured active contours end-to-end