Comparison of Deep-Learning-Based Segmentation Models: Using Top View Person Images

Image segmentation is considered as a key research topic in the area of computer vision. It is pivotal in a broad range of real-life applications. Recently, the emergence of deep learning drives significant advancement in image segmentation; the developed systems are now capable of recognizing, segmenting, and classifying objects of specific interest in images. Generally, most of these techniques primarily focused on the asymmetric field of view or frontal view objects. This work explores widely used deep learning-based models for person segmentation using top view data set. The first model employed in this work is Fully Convolutional Neural Network (FCN) with Resnet-101 architecture. The network consists of a set of max-pooling and convolution layers to identify pixel-wise class labels and prediction of the mask. The second model is based on FCN called U-Net with Encoder-Decoder architecture. The encoder is mainly comprised of a contracting path, also called an encoder, which captures the context in the image and symmetric expanding path called decoder to enable accurate location. The third model used for top view person segmentation is a DeepLabV3 model also with encoder-decoder architecture. The encoder consists of trained Convolutional Neural Network (CNN) to encode feature maps of the input image. The decoder is used for up-sampling and reconstruction of output using important information extracted by the encoder. All segmentation models are firstly tested using pre-trained models (trained on frontal view data set). To improve the performance, these models are further trained using person data set captured from a top view. The output of all models consists of a segmented person in the top view images. The experimental results reveal the effectiveness and performance of segmentation models by achieving $IoU$ of 83%, 84%, and 86% and $mIoU$ of 80% 82% and 84% for FCN, U-Net, and DeepLabv3 respectively. Furthermore, the discussion is provided for output results with possible future guidelines.