Research map / Vision-language and multimodal / Human and embodied modeling
When Visual Grounding Meets Gigapixel-level Large-scale Scenes: Benchmark and Approach
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 GigaGrounding, a gigapixel-scale visual grounding benchmark built from PANDA-Video, and GlaZing, a glance-to-zoom-in two-stage grounding model that outperforms one-stage and two-stage baselines.
Problem
Existing visual grounding benchmarks use megapixel images with few objects, so models are not tested on gigapixel-level large-scale scenes with many objects, high resolution, scale variation, and multi-hop expressions.
Approach
A coarse-to-fine "glance-to-zoom-in" cascade: a grounding module localizes the target on a downsampled thumbnail, an adaptive cropping module extracts a high-resolution patch around the predicted region, and a zoomed-in grounding module refines the box.
Key results
- GlaZing reaches 63.9% test Precision@0.5 at 1536x1536 versus 49.7% for the best baseline ReSC (a 28.6% relative gain).
- All one-stage and two-stage baselines score below 50% at 1536x1536 while human testers score 62% to 84%.
- MAttNet with ground-truth proposals reaches only 6.8% test accuracy, barely above random guessing at 6.2%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use glance-to-zoom-in grounding with adaptive cropping instead of ReSC | GigaGrounding grounding · GigaGrounding test, 1536×1536 | Precision@0.5: 49.7% -> 63.9% | Helped |
| Increase input resolution from 640×640 to 1536×1536 | ReSC · GigaGrounding test | Precision@0.5: 29.0% -> 49.7% | Helped |
| Initialize with RefCOCO weights | TransVG · GigaGrounding test, 640×640 | Precision@0.5: 5.1% -> 11.6% | Helped |
| Initialize with RefCOCO weights | RefTR · GigaGrounding test, 640×640 | Precision@0.5: 0.0% -> 21.8% | Helped |
| Initialize with RefCOCO weights | SeqTR · GigaGrounding test, 640×640 | Precision@0.5: 5.3% -> 15.9% | Helped |
| Use 8-times downsampled images in ACM and ZGM (ablation) | GlaZing · GigaGrounding test | Precision@0.5: 66.5% | Helped |
Surprising findings
- Two-stage MAttNet barely beats random guessing even when given ground-truth object boxes as proposals, which the authors attribute to frozen feature extractors.
- Pretraining on RefCOCO helps baselines on GigaGrounding despite large domain differences, but the gains are modest in absolute terms.
Limitations
- Glance and zoom modules are trained separately, not end-to-end.
- Only one glance-zoom iteration is explored; multiple iterations are left to future work.
- Privacy risk: individuals may remain identifiable from clothing or body shape despite face anonymization.
Experimental setup
Datasets: GigaGrounding RefCOCO
Benchmarks: GigaGrounding
Metrics: Precision@0.5 inference time GFLOPs
Baselines: MAttNet ReSC TransVG RefTR SeqTR QRNet SimREC
Builds on: ReSC TransVG PANDA
Closest papers by meaning
- Your Large Vision-Language Model Only Needs A Few Attention Heads For Visual Grounding (CVPR 2025 · similarity 0.89)
- From Objects to Anywhere: A Holistic Benchmark for Multi-level Visual Grounding in 3D Scenes (NeurIPS 2025 · similarity 0.89)
- GeM-VG: Towards Generalized Multi-image Visual Grounding with Multimodal Large Language Models (AAAI 2026 · similarity 0.88)
- GroundingME: Exposing the Visual Grounding Gap in MLLMs through Multi-Dimensional Evaluation (CVPR 2026 · similarity 0.88)
- GroundVLP: Harnessing Zero-Shot Visual Grounding from Vision-Language Pre-training and Open-Vocabulary Object Detection (AAAI 2024 · similarity 0.88)
- Connecting the Dots: Training-Free Visual Grounding via Agentic Reasoning (AAAI 2026 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.