MIDS Capstone Project Spring 2022

Where’s Waldo? Instance Level Image Retrieval

Instance level retrieval is the process of using a query image to search a database of images that matches the object. This process is often accomplished by extracting features from the images and then using these features in a retrieval process to retrieve images with similar features to the query image. Reranking then improves the retrieved images by reranking the images using techniques such as Geometric verification and Reranking Transformers. In this work we improve upon the current best practice of the joint optimization of feature extraction and reranking that uses Reranking Transformers as well as Geometric Verification to rerank the images to improve the accuracy of the images retrieved. We also work to create a prototype process which allows for a user to test their own image on the image retrieval model and to see the top images retrieved for their query image. Lastly, we introduced a new dataset that reapply’s the work we completed to see how adaptable this process is to other datasets. (Research in this paper is based off of the work done in this paper https://arxiv.org/pdf/2103.12236.pdf)

More Information

Last updated:

April 12, 2022