Lecture 1 - Course Introduction
Computer vision overview, Historical context, Course logistics
AdaBoost
Feature Based Object Recognition
- “Sift” & Object Recognition
- Problem: To map an entire object (stop sign) to another similar object is incredibly hard - due to camera angles, lighting, occlusion, viewpoint, intrinsic variation.
- However, there tend to remain features that remain “diagnostic and invariant to changes”
- Identifying critical features on an object, and match features to another object
- Using the same building block (diagnostic features in images), we have made another step forward as a field and started to recognize holistic scenes.
- Spatial Pyramid Matching
- Features in the images that can give us clues as to which type of thing it is
- Takes features from diff parts of images in diff resolutions and puts them together, then SVM is done on top of that
- Spatial Pyramid Matching
PASCAL Dataset
- Image recognition benchmark dataset
- Performance increases steadily over the years
Overfitting
- Overfitting happens very fast with not enough training data
- Solution: ImageNet - largest possible dataset of pictures, the world of objects, and use it for benchmarking and training
Deformable Part Model