Understanding Geometry

[NicolaBernini]

Overview

• Geometry is a fundamental element of Computer Vision
• In literature it is often called in different ways, including: Depth, Structure, ...
• Geometry is inherently related to Camera Motion
• The underlying Theory to deal with Geometry and Motion in Computer Vision is Epipolar Geometry

[NicolaBernini]

Geometry vs Semantic

Overview

• Computer Vision deals with both Geometry related and Semantic related Tasks
• Computer Vision has been positively impacted by the Deep Learning Revolution but so far it mostly regarded the Semantic intensive Tasks: Detection, Segmentation, ...
• Geometry related Tasks are still mostly solved with "traditional" (non Deep Learning based) methods but why

Possible reasons

• Semantic is discrete and subjective

• Dominant (the one that's providing best results) approach so far in DL is Supervised Learning which injects external (supervisor specific) semantic via supervision

• Geometry is a continuous quantity and it's objective

• Not the typical Supervision Signal provided by humans, nevertheless it is possible to obtain from other algorithms

  • E.g. PoseNet has been trained to regress 6DOF Pose from Mono Images (i.e. solving also Scale Ambiguity) with Supervised Learning using Structure from Motion as Supervision Signal

• Could be best suited for other learning paradigms like Unsupervised Learning