Contrastive pre-training using large, noisy image-text datasets has become popular for building general vision representations. These models align global image and text features in a shared space through similar and dissimilar pairs, excelling in tasks like image classification and retrieval. However, they need help with fine-grained tasks such as localization and spatial relationships. Recent efforts…

In the challenging fight against illegal poaching and human trafficking, researchers from Washington University in St. Louis’s McKelvey School of Engineering have devised a smart solution to enhance geospatial exploration. The problem at hand is how to efficiently search large areas to find and stop such activities. The current methods for local searches are limited…

There are two major challenges in visual representation learning: the computational inefficiency of Vision Transformers (ViTs) and the limited capacity of Convolutional Neural Networks (CNNs) to capture global contextual information. ViTs suffer from quadratic computational complexity while excelling in fitting capabilities and international receptive field. On the other hand, CNNs offer scalability and linear complexity…