Semi-Supervised Head Detection for Low Resolution Images

Sammanfattning: Object detection is a widely researched and applied field in computer vision. Deep learning models have successfully been used for object detection over the years. The performance of State of the art (SOTA) object detection deep learning models is dependent on the number of labeled images. Semi-Supervised Learning (SSL) methods in deep learning use both labeled and unlabeled images for training. This approach of training a model is gaining momentum as it saves the cost and time of labeling. In this project, a head detector for the football domain is trained using a SSL framework called soft teacher. An object detector trained on the COCO dataset [22], is adapted to the football domain to detect heads using the soft teacher [41] framework. The results of this model are compared to the supervised baseline and it was observed that the model trained with SSL framework adapts better to the football domain. For the experiments conducted with different percentages of labeled images, the 50 percent labeled setting had increase in performance by 9 percent compared to the 10 percent setting.