Ultra high resolution remote sensing vqa improves with self distillation
RS-OPSD: Reliable Privileged On-Policy-Self-Distillation for Ultra-High-Resolution Remote Sensing VQA
Computer Vision and Pattern Recognition
Summary
Ultra-high-resolution remote sensing images are very large and contain tiny details that must be noticed to answer questions about them. The paper's authors built a special training method called RS-OPSD that helps a model learn to focus on important small parts without extra searching later. They also created a new dataset to help the model identify evidence regions relevant to questions. This approach improves accuracy without slowing down the model, outperforming similarly sized competitors and even beating some much larger models.
What this means in practice
- •For geospatial analysts: Create faster and more accurate systems to answer questions about large satellite images without needing extra image searches at runtime.
- •For environmental monitoring teams: Build efficient VQA tools to analyze ultra-high-resolution images for detecting small-scale changes or objects relevant to environmental studies.
- •For security surveillance operators: Deploy models that rapidly interpret large-scale aerial images to identify critical regions based on user questions without costly real-time zooming or tool assistance.$Commercial implications: Enables commercial software for real-time analysis of high-res remote sensing images, improving surveillance capabilities.
Authors
Chengjie Jiang, Yunqi Zhou, Jiafeng Yan, Sihang Zhao, Chun Yuan, Jing Li
Abstract
Ultra-high-resolution (UHR) remote sensing visual question answering (VQA) requires models to resolve small visual evidence within extremely large images. Existing approaches typically rely on token pruning, visual search, or tool-augmented reasoning at inference time. We instead investigate whether the benefit of zoom-in visual privilege can be internalized into the model. We introduce RS-OPSD, a reliable privileged on-policy self-distillation (OPSD) framework for UHR remote sensing VQA. To provide high-quality privileged information with explicit question-relevant evidence, we construct GeoEvidence-6K, containing 6,750 VQA samples across seven task categories with evidence-region annotations, and develop Human Feedback-Guided Skill Refinement (HF-SR) for scalable annotation. To address context loss from tight crops and conflicting signals from imperfect teachers, RS-OPSD introduces Context-Preserving Visual Privilege (CPVP) and Correctness-Aligned Distillation (CAD). Without any additional visual search or tool calls at inference time, RS-OPSD achieves state-of-the-art (SOTA) performance on XLRS-Bench, MME-RealWorld-RS, and LRS-VQA, outperforming pervious SOTA models of comparable scale by an average of 4.0 percentage points. Moreover, our 2B variant, RS-OPD-Lite, surpasses most 8B-scale models while achieving the fastest measured inference speed. Our Code, GeoEvidence-6K, and the model weights for RS-OPSD and RS-OPD-Lite are publicly available.