The folder contains code for the pixel-wise classification of a road-scene image using different models like FCN32, FCN16, FCN8, Convolution-Deconvolution Model. The whole code is developed using TensorFlow library in python. This also contains unpooling operation in TensorFlow.
Real-time pixel wise segmentation of Road scene images using Deconvolution network with bispline upsampling
Semantic segmentation is a pixel-wise classification of an image where each pixel belongs to one of the classes like car, road, pedestrians etc in case of road scene. Real-time semantic segmentation is an active topic of research and very crucial for self-driving cars. Convolution - Deconvolution architecture has been used to increase the accuracy of the task but it requires to store indices from pooling operation which results in higher latency. We experimented different ways to do deconvolution process which requires less memory and hence less inference time. This included replacing the unpooling method by bi-spline upsampling and modifying deconvolution network to reduce the parameters involved.
- Tensorflow
- Numpy
- Matplotlib