Note The study investigates whether computer vision foundation models (e.g., DINO [9], OpenCLIP [22, 33]) exhibit similarities to the human visual system (HVS) in low-level perceptual tasks, such as contrast detection, masking, and constancy. OpenCLIP and DINOv2’s partial success suggests language-supervised and self-supervised learning may implicitly model some HVS features.

Note Bridge the resolution gap: - Inputs: low-resolution Sentinel-2 imagery (10 m). - Outputs: high-resolution (50 cm) building and road segmentation, counting, and height estimation. Effective teacher-student paradigm, a teacher model is trained on high-res data and the student model use pseudo-labels from teacher model

Note Reframes geo-localization as an image-to-GPS retrieval problem: - Image Encoder: Uses a CLIP-based Vision Transformer (ViT) - Location Encoder: The multi-scale approach with Random Fourier Features at various frequencies effectively captures both coarse- and fine-grained location information. - Contrastive Learning Objective

Note - Genuinely Unsupervised Vision-Language Alignment without text annotations, by using Ground Imagery as an intermediary - Two encoders: image-level encoder and pixel level encoder, and image-to-image global contrastive loss and pixel-to-image contrastive loss (sparse pixel alignment)