SURABHI: Self-Training Using Rectified Annotations-Based Hard Instances for Eidetic Cattle Recognition.

We propose a self-training scheme, SURABHI, that trains deep-learning keypoint detection models on machine-annotated instances, together with the methodology to generate those instances. SURABHI aims to improve the keypoint detection accuracy not by altering the structure of a deep-learning-based keypoint detector model but by generating highly effective training instances. The machine-annotated instances used in SURABHI are hard instances-instances that require a rectifier to correct the keypoints misplaced by the keypoint detection model. We engineer this scheme for the task of predicting keypoints of cattle from the top, in conjunction with our Eidetic Cattle Recognition System, which is dependent on accurate prediction of keypoints for predicting the correct cow ID. We show that the final cow ID prediction accuracy on previously unseen cows also improves significantly after applying SURABHI to a deep-learning detection model with high capacity, especially when available training data are minimal. SURABHI helps us achieve a top-6 cow recognition accuracy of 91.89% on a dataset of cow videos. Using SURABHI on this dataset also improves the number of cow instances with correct identification by 22% over the baseline result from fully supervised training.