Gym Equipment Visualization: an Object Detection Approach using Deep Learning on Mobile

In the realm of vision using mobile devices and image-processing networks, object detection presents a formidable challenge. Conducted entirely within Technogym S.p.A., this thesis embarked on a comprehensive exploration facilitated by the company's research-oriented vision. The project, granted expansive scope by Technogym's commitment to innovation, entailed a meticulous process. Beginning with an extensive review of tools and state-of-the-art detection products as well as company possible use cases for developing AI in apps, the research progressed to establish a tailored pipeline. This pipeline encompassed image collection, augmentation through in-house software, and iterative development of a model adept at processing the surrounding environment for gym equipment recognition. While the underlying neural network leveraged high-level implementations rather than custom-built architectures, numerous iterations (named executions) ensured to attain satisfactory recognition (approximately 70%-80% average precision). This achievement, though sufficient for the present, laid the groundwork for subsequent dedicated advancements. A significant portion of the thesis focused on constructing an iOS mobile app capable of real-time processing using the developed models. This not only provided the entire team with a practical benchmark for comparisons but also facilitated concrete demonstrations. The project's success is emphasized by its adherence to deadlines, enabling the delivery of tangible results promptly.