README

TODO Lists

• FCN
• PSPNet
• Deeplab v1
• Deeplab v2
• Deeplab v3
• Deeplab v3+
• HRNet
• U-net (Pix2Pix)
• Pix2PixHD
• Deformable Convolution v1
• Deformable Convolution v2
• BiSeNet

Summary

FCN-Fully Connected Network

VGG Backbones

32x down sampling

Using 1/8, 1/16, 1/32 sizes of feature maps from 3th, 4th, 5th downsampling of VGG feature map

Upsampling method: Transposed Convolution k=3, s=2, p=1, d=1 out_padding=1

BCEWithLogitsLoss

PSPNet - Pyramid Sence Parsing Net

Resnet34 Backbone

8x downsampling

using pooling window size of [1,2,3,6]

do AdaptiveAvgPool on last feature map

3 times (2**3=8) upsampling

Auxiliary loss & BCEWithLogitsLoss

Deeplab v1

Resnet34 Backbone

Stem 4x down sample

Res Layers 2x downsample

Total 8x downsample

Using Atrous Convolution/Dilated Convolution

CRF Conditional Random Field (Wasted in v3)

Deeplab v2

add ASPP Atrous Spatial Pyramid Pooling: using multiple Atrous Convolution replace pooling and concatresults

Deeplab v3

using multi-grids in res layer5 for de-griding

modifications on ASPP

Deeplab v3+

add a decoder to upsample results to the original size of input image

Upsample method: bilinear

Pix2PixHD

Too much

See https://github.com/NVIDIA/pix2pixHD

BiSeNet

Spatial Path (ResNet) + Attention Refinement Module (ARM) => Context Path

Feature Fusion Module (FFM)