Telecom Equipment Segmentation and Detection Using Drone Images

Abstract: An automated AI solution for out-door Telecom equipment segmentation is beneficial to most of the workflow for site survey and engineering performed by human. AI solutions that perform segmentation tasks are today trained with supervised learning which requires manually labeled images. However, labeling images is both time consuming and expensive, which makes semisupervised learning attractive where unlabeled data is used to further improve the performance of models. To determine if semi-supervised learning can be used to improve the performance of instance segmentation, the effectiveness of a semi-supervised learning approach called FixMatch was tested for instance segmentation using a custom dataset. The dataset contains 590 labeled and 1000 unlabeled drone-captured images of Telecom equipment. An extension was made to FixMatch where the predicted bounding boxes and masks are augmented like the images, which makes it possible to use FixMatch for instance segmentation. The extension was evaluated with mean Average Precision (mAP) but only achieved 1 point higher mAP than without using the extension. The small improvement in performance shows that this semi-supervised approach is not suitable for instance segmentation of Telecom equipment where the amount of unlabeled data is twice the labeled.