one-shot logo detection in the wild
Name: Meng Song
Grad Year: 2022
Despite significant advances in computer vision catalyzed by convolutional neural networks, most current models require millions of labeled examples to achieve state-of-the-art performance. However, manual labeling is not scalable for large-scale, long-tailed data, and we know that humans are good at learning new concepts with little supervision. In this regard, the research community is actively working on problems such as zero-shot, one-shot learning and transfer learning to generalize recognition models. However, few work have focused on investigating these problems in the context of object detection. In this project, we are aiming to solve the one-shot logo detection problem in the wild, where given an unseen clean logo, the model should correctly localize and identify it in any real image without re-training the model. To achieve this goal, we propose a novel framework which is different from traditional detection paradigms. Instead of reducing object detection to a classification and regression task, we formulate the detection problem as a feature embedding and verification process. As an important contribution of this framework, the verification network tries to find the best match from all possible pairs retrieved from the embedding space. To complement the feature embedding step, the network models the geometric transformation between the query patches and the standard logos. Experimental results demonstrate that our methods have achieved high performances on the unseen test logos, and the overall process can easily be improved by refining each individual module.
Industry Application Area(s)
Software, Analytics | Computer Vision, Artificial Intelligence