Neural network comparison - orthoseg/orthoseg GitHub Wiki

I first looked around to several sources to find the most promising neural networks to use for image segmentation.

Interesting pages I found in this search are the following:

• A very interesting comparison of the performance of different DNNs: Benchmark Analysis of Representative Deep Neural Network Architectures. Not specific to image segmentation though, rather object classification, but interesting to choose a good backbone DNN for the segmentation.
• This is an overview of the accuracy and network size of the different pretrained neural networks that are available in keras: Keras models.
• paperswithcode.com is also an interesting website with an huge overview of accuracy results achieved using AI, also in the domain of computer vision. Some examples:
  • Best performing image classifications on the imagenet dataset
  • Best performing image segmentations on the PASCAL VOC 2012 dataset

Based on the tests I did in 2018-2019, inceptionresnetv2 gives/gave the best results as a backbone, so this is the default in orthoseg.