IMPLEMENTATION OF YOLOV5 AND MASK R-CNN OBJECT DETECTION MODEL IN REAL TIME MOBILE APPLICATION FOR LOW VISION USERS

• Main Author: Anindya Riyadi, Inka
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/69191
• Institution: Institut Teknologi Bandung

Low vision is partial vision loss that can’t be corrected with glasses, contacts or surgery. It is not blindness who experience total vision loss. People who have low vision usually have blurry vision, therefore it is harder for them to do every day activities. Based on that problem statement, one of the latest applications of deep learning that can help low vision is object detection. However, several published mobile applications that helps low vision does not have capability to detect object by real time and none of it supports Bahasa Indonesia. Unfortunately, this caused the application can only be used for users who understands English. This research covers the implementation of Mask R-CNN and YOLOv5 using transfer learning method. Both models are trained but only one of the models will be implemented on the mobile application. The research will conduct experiment using various optimizer including SGD, Adam, AdamW when training the models to find which optimizer helps model to have the best performance. The result of the experiment is compared and analyzed in server. YOLOv5s – SGD have the highest mAP score with 0.39401 and YOLOv5s – AdamW have the fastest inference time. Both model with optimizer is further compared and analyzed in mobile. YOLOv5s – SGD inference time (930 ms) is faster than YOLOv5s – AdamW (3037 ms). Therefore, YOLOv5s – SGD is chosen as the best performing model and implemented on the mobile application.