hamidriasat/ddrnets

Deep Dual-resolution Networks for Real-time and Accurate Semantic Segmentation of Road Scenes

DDRNets

Deep Dual-resolution Networks for Real-time and Accurate Semantic Segmentation of Road Scenes

Hit star ⭐ if you find my work useful.

This repository contain implementation of DDRNet_23_slim in tensorflow/keras.

Comments are added for better understanding. Check below model images for better understanding of code.

Code Files

• ddrnet_23_slim.py → ddrnet_23_slim model code

Architecture

Dependies:
Tensorflow 2.0 or later

Licensed under MIT License

ddrnets's Issues

Training on cityscapes dataset

Hi

Thanks for providing your code. I want to ask about the performance of this network on cityscapes dataset. As in the paper the author report around 77% mIoU on cityscapes dataset. But I am using your code and I am getting mIoU around 55% on validation set of cityscapes dataset. i am training it on 680X680 image resolution. Have you trained your code on cityscapes dataset?

Operands could not be broadcast together with shapes

Operands could not be broadcast together with shapes (512, 1024, 64) (128, 256, 64)

Low to High

x_temp = layers.Activation("relu")(layers_inside[2])
x_temp = layers.Conv2D(highres_planes, kernel_size=(1,1), use_bias=False)(x_temp)
x_temp = layers.BatchNormalization()(x_temp)
x_temp = tf.image.resize(x_temp, (height_output, width_output)) # 1/16 -> 1/8
x_ = layers.Add()([x_, x_temp]) # next high branch input, 1/8

Pre-trained Weights

Hi @hamidriasat!

Do you have any pre-trained weights to share?

Thanks