CNN¶
- Models
- LeNet-5
- AlexNet
- VGG (ICLR 2014)
- NIN (ICLR 2014)
- Inception
- STN (CVPR 2015)
- U-Net (MICCAI 2015)
- ResNet (CVPR 2016)
- Deconvolution
- PixelShuffle (CVPR 2016)
- Fusing global feature
- CAM (CVPR 2016)
- DenseNet (CVPR 2017)
- CapsuleNet (2017)
- Deformable Convolution
- Deep Layer Aggregation, DLA (CVPR 2018)
- CSPNet (2019)
- Light-weights models
- Multi-Scale Feature Representations
- Visualization
- Image-to-Image Translation
- Optical character recognition, OCR
- Image Retrieval
- Object Detection
3D¶
Concept¶
Image Convolution¶
Weighting Sharing¶
Pooling¶
- downscale
- small translation-invariance
Downscale¶
- resize
- stride
- pooling
- max pooling
- avg pooling
Stride vs pooling¶
stride | pooling | |
---|---|---|
computation | less | 4 times more computation+pooling operate |
propagation | bad | better |
Upscale / Up-sampling¶
used for 1. reverse downscale, 2. increase resolution
usually required for pix 2 pix application, e.g. segmentation, super-resolution
Methods¶
- resize
resampling and interpolation - deconvolution Deconvnet
- PixelShuffle