Rock Climbing Handhold Classifier

Indoor rock climbing can be a difficult sport, with the many different types of rock wall holds making it difficult to gauge the difficulty of a route. This project presents a tool that allows climbers to plan their routes more effectively by viewing detailed images of the wall they plan to climb. The tool takes an image of a rock wall and labels all the holds detected within the image, as well as their types (jug, crimp, sloper, edge, pinch, pocket, foothold, or unclassified handhold). This is accomplished by creating a test dataset of manually labeled images and feeding them into a machine learning network based on the YOLO v7 architecture.

Ultimately YOLOv7 proved to be ineffective for the task at hand. While it is computationally fast, one of its greatest weaknesses is identifying similar objects in large clusters. As such, the model often failed to classify a single handhold and produced bounding boxes arbitrarily across the images. When trained with the proprietary online tool Roboflow much more accurate results were produced, with the primary issue being some types of holds being overfit. Potential future work involves testing the project approach with a different network such as RetinaNet, and finding more efficient ways to collect and label data.