Quality Evaluation of Image Segmentation in Mobile Augmented Reality

Mobile Augmented Reality (AR) facilitates a seamless interactive experience between actual and virtual environments. AR employs segmented images for various purposes such as object recognition, occlusion boundary estimation, and foreground-background separation. However, evaluating the quality of segmented images in mobile AR is challenging due to the limited resources of mobile devices. Existing solutions employ neural networks with many layers, making it difficult to deploy them on mobile devices. To address this issue, we propose techniques to modify the inputs so that we can reduce the number of layers in the neural network, making it possible to deploy for mobile devices. This idea is incorporated into our proposed SegQNet. It utilizes deep learning techniques based on convolutional neural networks (CNNs) to evaluate the quality of overlaid segmentation in mobile AR. SegQNet achieves high accuracy without the need for ground-truth images or expensive computations. Our experiments on Android smartphones demonstrate that SegQNet outperforms two state-of-the-art methods without incurring significant overhead.