The quality of visual features is crucial for a wide range of computer vision topics, e.g., scene classification, object recognition, and object detection, which are very popular in recent computer vision venues. All these image classification tasks have traditionally relied on hand-crafted features to try to capture the essence of different visual patterns. Fundamentally, a long-term goal in AI research is to build intelligent systems that can automatically learn meaningful feature representations from a massive amount of image data. We believe a comprehensive coverage of the latest advances on image feature learning will be of broad interest to ECCV attendees.
The primary objective of this tutorial is to introduce a paradigm of feature learning from unlabeled images, with an emphasis on applications to supervised image classification. We provide a comprehensive coverage of recently developed algorithms for learning powerful sparse nonlinear features, and showcase their superior performance on a number of challenging image classification benchmarks, including Caltech101, PASCAL, and the recent large-scale problem ImageNet. Furthermore, we describe deep learning and a variety of deep learning algorithms, which learn rich feature hierarchies from unlabeled data and can capture complex invariance in visual patterns.
1. Introduction
where do we get low-level representation from?
2. State-of-the-art Image Classification Methods
(1) Features
(2) Discriminative Methods
a. bag of words
Issues:
Spatial information is missed.
b. Spatial Pyramid Pooling
(3) Generative Model
3. Image Classification Using Sparse Coding
V1中的处理和Gabor小波变换类似,做边缘检测。
Sparse Coding大致意思是寻找一组基向量,使得所有input数据可以用这组基向量线性表示出来,而系数大部分为0,因此称为稀疏。
这个方法假设边缘是一个场景最基本的元素,用这种方法可以得到一个比像素更简洁的,更高层的表示。
主要步骤:
但是这样依然不如SIFT,有三个方法改善:
与SIFT结合就是把输入数据变为SIFT descriptors。
与K-means方法比较,发现sparse coding就是K-means的一种soft版本,任何时候需要用k-means得到字典的时候,用sparse coding都会提升效果。
4. Advanced Topics on Image Classification Using Sparse Coding
(1) Why sparse coding helps classification?
A "topic model" view to sparse coding:
A geometric view to sparse coding:
接下来讲述了SC在MNIST上的实验数据,当SC得到最小误差的时候,学习到的基向量就像数字,启发:研究数据中的几何特征可能对分类有帮助。
Local Coordinate Coding
Applications:
(2) Recent Advances in Sparse Coding for Image Classification
5. Learning Feature Hierachies and Deep Learning