Discovering and Analyzing Visual Structures

Introduction

The goal of this project is to shift the dominant paradigm of learning-based computer vision: instead of systems attempting to replace human interpretation by providing predictions, we will develop approaches to assist experts in identifying and analyzing patterns.

Background

Indeed, while the success of deep learning on visual data is undeniable, applications are often limited to the supervised learning scenario where the algorithm tries to infer a label for a new image based on the annotations made by experts in a reference dataset. In contrast, we will take as input images without any annotation, automatically identify consistent patterns, and model their variation and evolution, so that an expert can more easily analyze them.

Concept of Visual Structures

I will introduce and develop the concept of visual structures. Their key features will be:

• Interpretability, in terms of correspondences, deformations, or properties of the observed images.
• Ability to incorporate prior knowledge about the data and expert feedback.

I propose two complementary approaches to formally define and identify visual structures:

1. One based on analyzing correspondences.
2. The other on learning interpretable image models.

Application Domains

We will develop visual structures in two domains in which breakthrough progress will open up new scientific discoveries:

1. Historical documents.
2. Earth imagery.

For example, from temporal series of multispectral Earth images, we will identify types of moving objects, areas with different types of vegetation or constructions, and model the evolution of their characteristics, which may correspond to changes in their activity or life cycle.

Conclusion

Ultimately, experts will still be needed to select relevant visual structures and perform analysis, but DISCOVER will revolutionize their work, trivializing tedious annotation tasks and even allowing them to work on issues they would have been hard-pressed to identify in the raw data.