/PIDNet

First implementation of PIDNet in Tensorflow/Keras.

Primary LanguagePythonMIT LicenseMIT

PIDNet Hits license

PWC PWC PWC

PIDNet: A Real-time Semantic Segmentation Network Inspired from PID Controller

Hit star ⭐ if you find my work useful.

This repository contain implementation of all three versions of PIDNet in Tensorflow/Keras.

Overview

overview-of-our-method
An overview of the basic architecture of our proposed Proportional-Integral-Derivative Network (PIDNet).

P, I and D branches are responsiable for detail preservation, context embedding and boundary detection, respectively.

Detailed Implementation

overview-of-our-method
Instantiation of the PIDNet for semantic segmentation.

For operation, "OP, N, C" means operation OP with stride of N and the No. output channel is C; Output: output size given input size of 1024; mxRB: m residual basic blocks; 2xRBB: 2 residual bottleneck blocks; OP1\OP2: OP1 is used for PIDNet-L while OP1 is applied in PIDNet-M and PIDNet-S. (m,n,C) are scheduled to be (2,3,32), (2,3,64) and (3,4,64) for PIDNet-S, PIDNet-M and PIDNet-L, respectively. 
Dependies:
Tensorflow 2.3 or later

Licensed under MIT License