Visual localization, the process of determining a camera’s exact 6-DoF pose within a known environment, is fundamental to applications such as autonomous vehicles, augmented reality, and robotics. Traditional methods like GNSS offer only coarse position estimates and are often unreliable in indoor or visually challenging scenarios, underscoring the need for alternative, high-precision approaches. This thesis addresses these challenges by focusing on the image retrieval component within the visual localization pipeline—a critical stage that narrows the search space for computationally expensive local feature matching and thereby enhances overall system efficiency and accuracy. The research systematically evaluates state-of-the-art image retrieval models, including NetVLAD, AP-GeM, and SALAD, under diverse environmental conditions such as varying illumination, seasonal changes, and significant viewpoint differences.