UCLA

This thesis aims to implement two computer vision approaches for pet owners to find their lost pets in a more effective manner. We first introduce the product design of this thesis in real-life conditions. Then we implemented YOLOv3 for face extraction, BN-Inception embedding with proxy anchor loss model for face recognition, and combinations of different embedding methods and classification models for face verification. For face recognition, when our approach returns the top 5 most likely matched pet images to the query image, the face recognition model could find 45.45% of all the matching gallery images. For face verification, by using EfficientNet as an image embedding model and SVM with Gaussian kernel as the classification model, our approach achieves test AU-ROC scores of 88.89%. Finally, we will conclude this case study and discuss further improvements. The codes and datasets used and processed in this thesis could be accessed on the Github repository: AliceLLLLLan/PetFinding.