Research map / Vision-language and multimodal / Multimodal LLMs

ZoomEarth: Active Perception for Ultra-High-Resolution Geospatial Vision-Language Tasks

CVPR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Instruction-tuned multimodal assistants · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces LRS-GRO, a semi-automatically annotated ultra-high-resolution remote sensing VQA benchmark with ROI boxes, and ZoomEarth, a Qwen2.5-VL-3B cropping-zooming active perception model trained with SFT then GRPO using a Region-Guided distance reward.

Problem

Ultra-high-resolution remote sensing VQA is hard for MLLMs: dynamic resolution and token pruning are passive and add redundant visual tokens, and existing tool-use datasets rely on OCR/detection that does not transfer to RS imagery.

Approach

Active perception via a model that first sees a downsampled global image, predicts a bounding box for the region of interest, crops it from the original high-resolution image, and re-answers; a Region-Guided reward gives dense distance-based credit when IoU is zero.

Key results

• ZoomEarth (Qwen2.5-VL-3B, 512px input) reaches 53.76 Avg. Acc and 34.39 APO IoU on LRS-GRO, versus 53.67 for InternVL3-8B and 50.17 for VLM-R3 (w/ tools).
• In zero-shot evaluation ZoomEarth reports 44.10 on MME-RealWorld-RS, 40.20 on XLRS-Bench and 38.61 on GeoLLaVA-8k, average 40.97, above InternVL3-8B at 38.43.
• Removing the Region-Guided reward drops Avg. Acc from 53.76 to 52.79 (0.97 points), more than removing the IoU reward (0.82 points).

Surprising findings

• After SFT alone, invoking the cropping tool lowered performance (42.80 to 41.30 overall), suggesting RL is needed for effective tool use.
• Increasing input resolution from 512 to 3333 px without cropping gave no gain (50.86 vs 50.92 Avg. Acc) and reduced speed.

Limitations

• Downstream tool integrations (cloud removal, denoising, segmentation, editing) are only demonstrated qualitatively in a training-free setting without quantitative evaluation.
• Annotations rely on GPT-4o generation followed by manual filtering, and the training set is small (about 3,500 samples).
• Cloud-covered inputs for the cloud removal demo were simulated with an image editing model because suitable data was unavailable.

Experimental setup

Datasets: LRS-GRO FAIR1M-1.0 GLH-Bridge STAR

Benchmarks: LRS-GRO MME-RealWorld-RS XLRS-Bench GeoLLaVA-8k

Metrics: accuracy APO IoU

Baselines: LLaVA-OV-1.5 IXC-2.5 InternVL3 GeoChat Qwen2.5-VL VLM-R3

Builds on: GRPO Qwen2.5-VL VLM-R3 LRS-VQA GeoChat

Closest papers by meaning

• GeoLLaVA-8K: Scaling Remote-Sensing Multimodal Large Language Models to 8K Resolution (NeurIPS 2025 · similarity 0.89)
• GeoViS: Geospatially Rewarded Visual Search for Remote Sensing Visual Grounding (CVPR 2026 · similarity 0.88)
• UHR-BAT: Budget-Aware Token Compression Vision-Language model for Ultra-High-Resolution Remote Sensing (ICML 2026 · similarity 0.88)
• SenseBench: A Benchmark for Remote Sensing Low-Level Visual Perception and Description in Large Vision-Language Models (NeurIPS 2026 · similarity 0.88)
• SkySense-O: Towards Open-World Remote Sensing Interpretation with Vision-Centric Visual-Language Modeling (CVPR 2025 · similarity 0.87)
• Towards Faithful Reasoning in Remote Sensing: A Perceptually-Grounded GeoSpatial Chain-of-Thought for Vision-Language Models (ICLR 2026 · similarity 0.87)