ridash2005

DRISHTI is a state-of-the-art Vision-Language Model (VLM) framework for Satellite Imagery. Built on Qwen3-VL and SAM3, it features a 180K+ sample dataset, Two-Stage LoRA/DPO training, and Pyramidal Tiling for high-accuracy counting and grounding. Achieved +40% BERT-BLEU in captioning and +21% VQA accuracy over baselines.