Scribd
Upload
15 views3 pages

Why Deeplab v3

Uploaded by

abdelrahmanalisaad011
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as DOCX, PDF, TXT or read online on Scribd
15 views3 pages

Why Deeplab v3

Uploaded by

abdelrahmanalisaad011
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as DOCX, PDF, TXT or read online on Scribd
You are on page 1/ 3

Why deepLab v3+ ?

Traditional DCNNs have some problems :

1 - Information Loss:

With Max pooling, there is a significant reduction in the resolution of the image, leading to a
loss of information

2 - Difficulty in Object Localization:

Due to the loss of information caused by Max


pooling, it becomes challenging to precisely locate or
localize objects in the image

Difficulty in Object Localization

3 - Multiscale Contextual Information:

Objects in an image can exist at different


scales. However, traditional Max pooling does not
consider the multiscale contextual information.

Multiscale Contextual Information


These limitations led to the development of alternative techniques, such as Atrous convolution and
Atrous Spatial Pyramid Pooling (ASPP), used in models like DeepLab, to overcome these issues and
improve the performance of semantic segmentation tasks.

Atrous Convolution:

Atrous convolution, also known as dilated convolution, is a variant of traditional convolution that allows
for the expansion of the receptive field without increasing the number of parameters. It is a
fundamental operation used in many advanced convolutional neural network (CNN) architectures,
including DeepLab.

Atrous Spatial Pyramid Pooling ( ASPP) :

ASPP uses in DeepLab, to capture multi-scale contextual information and improve the accuracy
of segmentation results. It consists of parallel atrous convolutions with different dilation rates
and a global pooling branch.
Here's how ASPP works:

1. Parallel Atrous Convolutions: ASPP includes multiple atrous convolutional layers with
different dilation rates.
2. Global Pooling Branch: ASPP also includes a global pooling branch that performs
pooling operations, such as average pooling or max pooling, on the feature maps. This
branch aggregates information across the entire spatial extent of the feature maps,
capturing global context information. The pooled feature representation is then used to
capture high-level contextual information.
3. Fusion: The outputs from the parallel atrous convolutional layers and the global pooling
branch are combined through concatenation or aggregation operations. This fusion step
brings together features captured at different scales and with different contextual
information.

You might also like