Supplements
Chap 01: Introduction to CV
- Why CV so important? over 85% of data on Internet is pixel-based, images or videos => Vision.
- Why is CV exploding?
- Sensors. All kinds of cameras everywhere. (smartphones, cars, drones, surveillances)
- Big Data. Dark matters on Internet. beyond human ability.
Chap 15: Convolutional Neural Networks
Chap 27: Segmentation & (Soft)Attention
- Segmentation
- semantic segmentation : 只输出不同的类别class
- instance segmentation: 即使是同一个class, 不同的个体instance, 也独立标出
- (Soft)Attention
- Discrete locations
- Continuous locations (Spatical Transformers)
multi-scale, upsampling, downsampling, deconvolution
-
FCN
-
Multi-scale
-
Refinement : iteratively
-
Upsampling
- Deconv = normal conv back-propagation pass, a bad name, sounds like
inverse of conv
, in fact, it meansconv transpose
, orupconvolution
,backward strided conv
or1/2 strided convolution
.