Research map / Vision-language and multimodal / Multimodal LLMs

See What We Cannot See: A Geo-guided Reasoning Benchmark for Object Counting under Adverse Earth Observation Conditions

CVPR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Spatial reasoning in VLMs

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 GROC, a 1.2M-point, 14K-image remote sensing object counting benchmark with paired clear/degraded scenes and three geo-modalities (land use, map, DSM), plus a data engine and a GPT-5-based geo-guided counting agent baseline.

Problem

Object counting in remote sensing fails when clouds, fog, shadows, or low light obscure visual cues, and existing benchmarks lack geo-modalities and paired clear/degraded images needed to evaluate geo-guided amodal reasoning.

Approach

A benchmark with aligned geo-modalities (land use, map, DSM) and synthetic controllable degradations (clouds, shadows, low light) applied consistently to RGB and NIR, together with a tool-using MLLM agent that selects expert counters and uses geo-priors when appearance fails.

Key results

• GROC Agent achieves the best adverse-condition MAE of 14.476 (weather) and 13.518 (illumination) overall adverse MAE 3.940 on clear test images vs 8.418 for DINO-X under clear scenes... (clear: 3.940 MAE, RMSE 7.560)
• On degraded scenes, Qwen3-VL-8B-Instruct reaches MAE 40.302 on clear degraded-split, far worse than specialized counters such as BL (4.968 clear MAE).
• Density models such as BL and FIDTM remain comparatively stable under adverse weather; DINO-X is more resilient under low light (MAE 13.924 vs 24.647-44.307 for others on illumination).

Surprising findings

• Generic MLLM Qwen3-VL-8B-Instruct performs far worse than specialized counting models even on clear scenes (MAE 11.330 average for car vs 5.131 for the agent).
• Gains of the GROC Agent remain limited under severe degradation where visual evidence is largely unavailable.

Limitations

• Degradations are synthetically generated rather than real adverse-condition captures, though realism validation is claimed in supplementary material.
• Agent gains are limited under severe conditions where visual evidence is largely absent.
• Only four object categories (car, truck, boat, airplane) are covered.

Experimental setup

Datasets: GROC

Benchmarks: GROC

Metrics: MAE RMSE

Baselines: MCNN CSRNet BL SFCN SCAR PSGCNet FIDTM GroundingDINO DINO-X Qwen3-VL-8B-Instruct

Builds on: DINO-X Perlin turbulence cloud synthesis C-3-Framework

Closest papers by meaning

• InstructSAM: A Training-free Framework for Instruction-Oriented Remote Sensing Object Recognition (NeurIPS 2025 · similarity 0.90)
• Towards Faithful Reasoning in Remote Sensing: A Perceptually-Grounded GeoSpatial Chain-of-Thought for Vision-Language Models (ICLR 2026 · similarity 0.89)
• TerraScope: Pixel-Grounded Visual Reasoning for Earth Observation (CVPR 2026 · similarity 0.88)
• EarthVQA: Towards Queryable Earth via Relational Reasoning-Based Remote Sensing Visual Question Answering (AAAI 2024 · similarity 0.88)
• GeoViS: Geospatially Rewarded Visual Search for Remote Sensing Visual Grounding (CVPR 2026 · similarity 0.87)
• RemoteReasoner: Towards Unifying Geospatial Reasoning Workflow (AAAI 2026 · similarity 0.87)