Research map / Detection and segmentation / Open-set and few-shot detection
DAVE - A Detect-and-Verify Paradigm for Low-Shot 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 DAVE, a low-shot counter that first generates high-recall detections from a density-based backbone and then verifies them via clustering of appearance features to remove outliers, improving both counting and detection across few-shot, zero-shot, and text-prompt settings.
Problem
Low-shot counting methods either give accurate total counts without object locations (density-based) or give locations with poorer counts (detection-based), and both overcount in scenes with other object classes due to false positives.
Approach
A two-stage detect-and-verify pipeline: a LOCA-based density model yields high-recall candidate boxes; a verification network extracts appearance features for each candidate and exemplars, spectral clustering keeps only clusters containing an exemplar, and the density map is masked to the kept detections to refine counts.
Key results
- DAVE reduces test MAE on FSC147 to 8.66 vs 10.79 for LOCA (few-shot density-based), about 20% relative improvement.
- On FSCD147 test, DAVE reaches 26.81 AP and 62.82 AP50, versus 22.66 AP and 50.57 AP50 for C-DETR.
- Box-count MAE of DAVE on FSC147 test is 10.45 vs 16.79 for C-DETR, a 38% improvement.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use detect-and-verify instead of density-only counting | DAVE on FSC147 test, few-shot · FSC147 test, few-shot | MAE: 10.79 -> 8.66 | Helped |
| use prompt-based cluster selection (ablation) | DAVE on FSC147mul · FSC147mul | MAE: 12.86 -> 6.46 | Helped |
| train LOCA on multi-class images (ablation) | LOCA · FSCD147mul | MAE: 21.28 -> 13.25 | Helped |
| train LOCA on multi-class images (ablation) | LOCA · FSC147 | MAE: 10.79 -> 12.63 | Hurt |
| add the verification feature projection network ϕ (ablation) | DAVE · FSCD147 | MAE: 9.41 -> 8.91 | Helped |
| add prototype correlation response tensor R̃ to box regression (ablation) | DAVE · FSCD147 | AP: 19.50 -> 24.19 | Helped |
| use FFM instead of concatenation (ablation) | DAVE · FSCD147 | AP: 18.49 -> 24.19 | Helped |
| use FFM instead of sum (ablation) | DAVE · FSCD147 | AP: 20.74 -> 24.19 | Helped |
Surprising findings
- Training LOCA on multi-class images reduces error on multi-class subset by 43% but increases error on FSC147 by 28%.
Limitations
- Count discrepancy is largest in extremely dense images with many small packed objects.
- Detection stage struggles in extreme density regions, limiting box-count accuracy.
Experimental setup
Datasets: FSC147 FSCD147 FSCD-LVIS FSCD-LVISuns
Benchmarks: FSC147 FSCD147 FSCD-LVIS FSCD-LVISuns
Metrics: MAE RMSE AP AP50
Baselines: LOCA CounTR SAFECount BMNet+ FamNet C-DETR CounTX CLIP-Count RCC RepRPN-C
Builds on: LOCA C-DETR CLIP FFM
Closest papers by meaning
- A Novel Unified Architecture for Low-Shot Counting by Detection and Segmentation (NeurIPS 2024 · similarity 0.92)
- Point Segment and Count: A Generalized Framework for Object Counting (CVPR 2024 · similarity 0.89)
- T2ICount: Enhancing Cross-modal Understanding for Zero-Shot Counting (CVPR 2025 · similarity 0.88)
- CountGD: Multi-Modal Open-World Counting (NeurIPS 2024 · similarity 0.87)
- OmniCount: Multi-label Object Counting with Semantic-Geometric Priors (AAAI 2025 · similarity 0.87)
- Single Domain Generalization for Few-Shot Counting via Universal Representation Matching (CVPR 2025 · similarity 0.87)
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.