流媒体技术课程笔记
Two learning manners
-
Bottom-up
stimulus-driven, mostly obtained from early features, task independent
-
Top-down
related to recognition processing, influenced by prior knowledge (e.g. tasks to be performed, feature distribution of target, context of visual scene, etc.)
Saliency
- Saliency Detection
Evaluation
Classification: ROC Curve and AUC
-
ROC
\[\text{TODO:}\ Alg.\] -
AOC: area under curve
- shuffled AUC (sAUC) [Zhang et al.]: eliminating edge effect
Salient Object Detection
Evaluation
Classification: Precision and Recall
- PR curve: precision, recall, and F-measure
SLIC Superpixels Segmentation
Application
- Manifold ranking 流形排序
- Absorbing Markov-chain
Intuition
受位置约束的颜色空间 K-means 分类
- uniform seeding
-
calculate distance, and label after argmin seed
\[D(k,i) = d_{lab}(k,i) + \frac{m}{S} d_{xy}(k,i)\] - enforce connectivity
- remove small isolated regions
Semi-Supervised Learning
- exploit prior knowledge from unlabeled data
Local and Global Consistency (LGC)
Assumption
- nearby samples are likely having same label
- samples on same structure (refered to as cluster / manifold) are likely having same label
Take assumption to math…
cost function / energy function (*proposed, may have other forms)
\[\begin{aligned} &E(F) = \underbrace{\sum_i(F_i - Y_i)^2}_\text{label fitness} + \underbrace{\frac{1}{2} \lambda \sum_{ij} w_{ij} (\frac{F_i}{\sqrt{d_{ii}}} - \frac{F_j}{\sqrt{d_{jj}}})^2}_\text{manifold smoothness} \\ \Rightarrow &F \sim (I - \frac{\lambda}{1+\lambda}D^{-\frac{1}{2}}WD^{-\frac{1}{2}})^{-1} Y \end{aligned}\]BP Network
- 若希望输出和原始输入一样,则为常见的自编码网络
- 正交矩阵,输入特征正交性较好,互不相关