Concepts
Name | Description | y y |
---|---|---|
Object Classification | At most one object | |
Object Localization | At most one object | y=⎛⎝⎜⎜⎜⎜⎜⎜⎜⎜⎜⎜⎜⎜⎜pcbxbybwbhc1c2c3⎞⎠⎟⎟⎟⎟⎟⎟⎟⎟⎟⎟⎟⎟⎟ y = ( p c b x b y b w b h c 1 c 2 c 3 ) |
Landmark Detection | At most one object | y=⎛⎝⎜⎜⎜⎜⎜⎜⎜⎜⎜pcl1xl1y⋮l64xl64y⎞⎠⎟⎟⎟⎟⎟⎟⎟⎟⎟ y = ( p c l 1 x l 1 y ⋮ l 64 x l 64 y ) |
Object Detection | Multiple objects |
Sliding Windows Object Detection Algorithm
- Input: closely cropped images by sliding window
- Crop images with larger window
Disadvantage
Complex computation
Convolutional Implement of Sliding Windows
Turn Full Connection (FC) into convolutional layer
Bounding Box Detection
Yolo (You only look once) Algorithm
IoU (Intersection over Union)
Non-max Supression (NMS)
Discard all boxes with low
pc
p
c
While remaining boxes exists:
1. Pick the box with the largest
pc
p
c
, output that as a prediction
2. Discard remaining boxes with IoU
≥0.5
≥
0.5
with the box output in previous step
Region Proposal
R-RNN: Segmentation Algorithm
Fast R-RNN
Faster R-RNN