Home page

Deconstructing Deep Learning + δeviations

Drop me an email | RSS feed link : Click
Format : Date | Title
  TL; DR

Total number of posts : 89


View My GitHub Profile

Go to index

LightSeg (only notes for now)

Reading time : ~3 mins

by Subhaditya Mukherjee

Paper notes for the paper

[36] LightSeg (only notes for now)

(P.S. This was done for an internship so is a bit more formal than usual)


TL;DR -> Faster Semantic Segmentation with a modified ASPP module(from the DeepLab paper) + MobileNetV2


Modules used

Atrous -> Dilated conv

- Decrease receptive field
- Upsample by adding 0s b/w two filter vals along spatial dimensions with a dilation factor
- Note for Pytorch -> Conv2d has dilation as a param ### ASPP
- Take feature map -> Add 4 parallel atrous convs with different rates ### Deeper Atrous Spatial Pooling(DASPP)
- Take ASPP
- Add 3x3 convs after 3x3 atrous convs ### Depthwise separable conv
- Replace normal convs
- Split input and output into channels
- Convolve pointwise
- Note for Pytorch -> Add groups = no of in_channels in Conv2d ### Residuals
- Take residual/skip connections 
  - H(X) = F(X) + X  ; H(X) is output; F(X) is residual  ; X is input feature map
- Long -> Across larger no of layers
- Short -> Smaller no of layers (as memory units)
- Fuse them both 
- Use 1x1 convs
- End up with richer features ### Encoder
- MobileNetV2 with output stride of 32 ### Decoder
- DeepLabv3+ decoder with ASPP modified to DASPP

Whats new

Misc training info

Results (mIOU)

Related posts:  FP16  AI Superpowers Kai Fu Lee  Digital Minimalism Cal Newport  More Deep Learning, Less Crying - A guide  Super resolution  Federated Learning  Taking Batchnorm For Granted  A murder mystery and Adversarial attack  Thank you and a rain check  Pruning