Campus-Level Geolocation and Building Instance Classification Using Transfer Learning

Visual geolocation, the process of determining the geographic location of an image based on its visual content, has become an increasingly important aspect of computer vision with diverse applications in navigation systems, augmented reality, security, urban planning, and environmental monitoring. This study investigates the application of deep convolutional neural networks (CNNs) for campus-level visual geolocation and building identification using a dataset of building images from the University of Salford. The dataset, consisting of photographs captured using a mobile phone camera from various angles and distances, presents challenges such as fixed and removable objects, variable lighting conditions, and dynamic elements. EfficientNetV2M and ResNet101 models were fine-tuned and evaluated for their performance in classifying the buildings. ResNet101 demonstrated superior performance with an overall precision of 0.89, recall of 0.86, and F1-score of 0.87, compared to EfficientNetV2M's precision of 0.82, recall of 0.81, and F1-score of 0.81. Despite ResNet101's higher overall metrics, EfficientNetV2M showed strong performance in recall, particularly for buildings like the Sports Center and Chapman.