In this thesis, we address the problem of sketch based image retrieval. We present a novel model that enables fine-grained scene-level retrieval. The proposed model is based on a reliable and straightforward approach, designed following the principles of a Siamese network. As an encoder, we use the modern convolutional architecture ConvNeXt. The traditional triplet loss function is replaced with the more advanced InfoNCE loss function. We evaluate the model on FSCOCO, the latest and largest available dataset for scene-level image retrieval. The proposed model outperforms the baseline model by nearly 30 percent. We also compare it with multimodal retrieval models, surpassing them by more than 20 percent. Additionally, we conduct evaluations that further confirm its robustness. Furthermore, we perform a user study in which we analyze instances of the model's failures and highlight challenges related to the dataset that the model encounters.
|
