- how do we find images in another image when there’s color diff, projection/angle diff, perspective diff, occlusion?
intuition
- find matching patches, check for enough patches
- identify patch
- learn a way to describe each patch
- need algorithm to match description between two patches
- matching large patches is challenging, so local image patches is better
- smaller patches are more likely to find an object even with occlusion
- articulation, intra-category variations
search
- How to choose key-points? We want to detect same points independently in both images - a repeatable detector
- How do we distinctively describe each patch such that we can reliably from a point recognize the corresponding one?
- descriptions should be robust/invariant to
- rotation and translation
- photometric transformations
- modeled as linear transformation scaling + offset
requirements summary
- existing feature detectors