Example YOLO output
https://www.youtube.com/watch?v=69Ii3HjUiTM

******************************************************************************

train with labelled data

*****************************************************************************
Previous techniques

Talk about
    k-Nearest Neighbors
    Decision Trees
    Logistic Regression

    Support Vector Machine
    Naive Bayes
    Random Forest
    Gradient Boosting

Effect of priors
Bayes law    
    
*****************************************************************************

What is an image?
W*H*3 -> huge vector of numbers
fully connected network with that input is too big.
translation invariance

What is a CNN?
convolve local operators that are applied all over the image
output is some number of features per location

What is pooling?
convolve local operators like max/min or sum/average
stride reduces resolution
goals
- better features
- less data

Purpose of RelU
nonlinear processing

Backend - ``flatten'' or stop convolving
What is the purpose of fully connected layers?
Global interactions (2 eyes, for example).
Could do multi-scale

Other tricks
normalization layers
skip connections
fixed weights?

what about rotation invariance?

****************************************************

1 step
YOLO

2 step
Propose regions
classify regions

******************

YOLO

Real time
Process grid cells
 many overlapping grids
 - random patch shapes (more than rectangle) of different sizes?
 - use some feature to guide patch selection
 non-maximul suppression to get rid of multiple predictions of the same thing
not as good as much slower algorithms Fast RCNN

How score?
Intersection Over Union
 IOU > 0.5 -> positive prediction
Average Precision - area under precision vs. recall curve
 Recall = total correct positive predictions made by model/total positive instances out there
 Precision = true positives/total predictions made by moddel
Cross validation
 
Has trouble with multiple objects in same grid cell.

*****************

These networks have 10s of millions of parameters/weights

Multilabel classification (multiple outputs)?

*****************

Segmentation

Bounding box -> actual borders

train with labelled data

Region growing, split, and merge - group pixels or small regions - local similarity
Edge-based methods - find borders - local differences
Threshold methods - classify pixels or small regions - mean shift
Watershed methods - start region growing at local max distance
Clustering methods - k-means - global similarity

encoder - decoder
convolution encode bottleneck decode expand
downsampling process upsampling
add skip connections

*********************************************************

Latent spaces

*********************************************************

LSTMs
recurrent networks