Hobby Projects

Image Segmentation with FCNs

My GitHub repo now contains a collection of codes includes a PyTorch implementation of the fully convolutional (FCN) version of Alexnet (see https://arxiv.org/abs/1411.4038), which detects objects in images. The code in this package has been written with one goal in mind: keeping it simple. The model is written only for 1 type of object to be detected (in addition to background), and all functions and classes are explicit; nothing is pre-trained. The goal is not to provide a powerful state-of-the-art model, but to make the basics clear.

This code was tested with the Penn-Fudan pedestrian dataset, as well as cat images from the COCO dataset. Even though the dataset is small, the results are reasonable. More accurate results can be obtained by using more elaborate neural network architechtures (e.g. transfer learning with a pre-trained ResNet-50).

(Images taken from the COCO dataset: https://cocodataset.org/)