Research map / Vision-language and multimodal / Human and embodied modeling
Exploring Contextual Attribute Density in Referring Expression Counting
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.
Proposes CAD-GD, which adds a U-shape contextual attribute density estimator and CAD-based attention and query initialization to GroundingDINO for referring expression counting.
Problem
Referring expression counting (REC) models over-count or under-count when fine-grained attributes must be aligned with the correct visual regions, especially under occlusion and scale variation.
Approach
Define contextual attribute density (CAD) as the information intensity of a fine-grained attribute in visual regions, estimate it with a U-shape estimator supervised by density maps built from text-image similarity features, and inject it into visual features via spatial/channel attention and into decoder query initialization.
Key results
- On REC-8K test with Swin-T, CAD-GD† reaches MAE 4.59 vs GroundingREC 6.50 (text says 33.2% relative error reduction on test).
- With Swin-B, CAD-GD achieves F1 0.75 on val and 0.76 on test, versus GroundingREC* F1 0.71 and 0.70.
- Zero-shot on FSC-147 test, CAD-GD obtains MAE 10.35 and RMSE 86.88, lower than other text-prompted methods such as CountGD (12.98 MAE).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add contextual attribute density modeling | GroundingREC with Swin-B · REC-8K val | MAE: 5.66 -> 4.83 | Helped |
| Add contextual attribute density modeling | GroundingREC with Swin-B · REC-8K test | MAE: 5.42 -> 4.94 | Helped |
| Use the density-map estimate to select the number of queries (ablation) | CAD-GD · REC-8K val | MAE: 4.83 -> 4.23 | Helped |
| Add contextual attribute density modeling | GroundingREC · FSC-147 val | MAE: 10.06 -> 9.30 | Helped |
| Add contextual attribute density modeling | GroundingREC · FSC-147 test | MAE: 10.12 -> 10.35 | Hurt |
| Add contextual attribute density modeling | CountGD · CARPK test | MAE: 3.83 -> 3.29 | Helped |
| Add the CAD Generate Module (ablation) | REC-8K ablation baseline · REC-8K val | MAE: 6.52 -> 6.17 | Helped |
| Add spatial attention (ablation) | CAD-GD ablation · REC-8K val | MAE: 6.17 -> 5.88 | Helped |
| Add channel attention (ablation) | CAD-GD ablation · REC-8K val | MAE: 5.88 -> 5.61 | Helped |
| Add text-based query initialization (ablation) | CAD-GD ablation · REC-8K val | MAE: 5.61 -> 5.67 | Hurt |
| Add CAD-based query initialization (ablation) | CAD-GD with text-based query initialization · REC-8K val | MAE: 5.67 -> 5.43 | Helped |
Surprising findings
- Adding text-only query initialization slightly lowered validation MAE in the ablation, which the authors attribute to reduced perception of the query image.
Limitations
- Localization recall does not improve over GroundingREC with Swin-T; the authors attribute gains mainly to precision.
- Density supervision relies on Gaussian density maps generated for REC-8K, which the original dataset lacks.
- Quality of CAD depends on the quality of visual features from the backbone, as the authors note for Swin-B.
Experimental setup
Datasets: REC-8K FSC-147 CARPK
Benchmarks: REC-8K FSC-147 CARPK
Metrics: MAE RMSE Precision Recall F1
Baselines: GroundingREC GroundingDINO ZSC CounTX CountGD CLIP-count VLCounter CACViT DAVE LOCA
Builds on: GroundingDINO GroundingREC CountGD
Closest papers by meaning
- Revisiting Counterfactual Problems in Referring Expression Comprehension (CVPR 2024 · similarity 0.88)
- Referring Expression Counting (CVPR 2024 · similarity 0.87)
- Rethinking Two-Stage Referring Expression Comprehension: A Novel Grounding and Segmentation Method Modulated by Point (AAAI 2024 · similarity 0.86)
- Unveiling Parts Beyond Objects: Towards Finer-Granularity Referring Expression Segmentation (CVPR 2024 · similarity 0.85)
- RESAnything: Attribute Prompting for Arbitrary Referring Segmentation (NeurIPS 2025 · similarity 0.85)
- Position Is All You Need: A Free Lunch Token Compression Strategy for MLLM-based Referring Expression Segmentation (ICML 2026 · similarity 0.84)
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.