Patch-wise image similarity search Searching for small regions in collections of large images
Examensarbete för masterexamen
In autonomous driving it is important that a neural network performs well even for examples that do not occur often in the dataset. One method to improve performance is to find and add examples similar to those the network struggles with, thereby increasing the training data available. Similarity search is an automatic method for searching large datasets for the most similar examples to some target. This thesis describes a similarity search algorithm for locating different sized objects in a large dataset of high-resolution images. The algorithms uses patches, which are small regions in an image, in order to enable precise searches for small objects. Each patch is embedded to 512 dimensional vectors with CLIP that captures the semantic meaning of the content in the patch. The main contribution of this thesis is a method to reduce the potential number of patches resulting from each image, by selecting the most visually interesting patches in each image. We evaluate the combined patch selection and similarity search on three classes of objects relevant for autonomous driving: ambulances, animals and a specific traffic sign marking an upcoming road narrowing, to measure the fraction of relevant patches retrieved. Further, we show that searching with the average vector representation of several images of the same object improves the result, while searching with a text string gives varying results depending on the object class.
thesis , nearest-neighbor search , deep learning , image retrieval