Research map / Detection and segmentation / Open-set and few-shot detection
A Novel Unified Architecture for Low-Shot Counting by Detection and Segmentation
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 GeCo, a single-stage low-shot counter built as an add-on to a frozen SAM backbone that performs counting, detection and segmentation, using dense object queries and a new Hungarian-matched detection loss.
Problem
Low-shot object counting by detection: existing detection-based counters overgeneralize prototypes (false positives), rely on a unit-Gaussian surrogate loss sensitive to annotation noise, and lag density-based counters in total count accuracy.
Approach
Generalize prototypes image-wide into location-specific dense object queries via non-parametric cross-attention (no spatial encoding), decode them into dense detections, and train with a loss that Hungarian-matches local maxima to ground-truth boxes (gIoU plus objectness targets for TP/FN/FP) instead of Gaussian heatmap targets.
Key results
- On FSCD147 test (three-shot), GeCo reaches 7.91 MAE and 54.28 RMSE, about 24% lower MAE than DAVE (10.45) and 39% lower than PSECO (13.05).
- On FSCD147 test, GeCo achieves 43.42 AP and 75.06 AP50, the best detection scores among compared methods.
- One-shot GeCo gets 8.10 test MAE versus 11.54 for DAVE one-shot, and zero-shot GeCo gets 13.30 test MAE versus 15.51 for DAVE.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace the proposed dense detection loss with standard unit-Gaussian loss (ablation) | GeCo · FSCD147 validation | MAE: 9.52 -> 12.79 | Hurt |
| Add SAM-HQ features during query unpacking (ablation) | GeCo · FSCD147 validation | MAE: 10.04 -> 9.52 | Helped |
| Add shape prototypes (ablation) | GeCo · FSCD147 validation | MAE: 9.97 -> 9.52 | Helped |
| Add bounding box refinement (ablation) | GeCo · FSCD147 validation | AP: 24.63 -> 33.51 | Helped |
| Use dense object query construction instead of standard self-attention (ablation) | GeCo · FSCD147 validation | MAE: 10.32 -> 9.52 | Helped |
| Use image features as queries rather than DETR-like object-query construction (ablation) | GeCo · FSCD147 validation | MAE: 11.45 -> 9.52 | Helped |
| Use GeCo rather than DAVE for few-shot counting | FSCD147 test, 3-shot · FSCD147 test | MAE: 10.45 -> 7.91 | Helped |
| Use GeCo rather than DAVE for few-shot counting | FSCD-LVIS unseen test, few-shot · FSCD-LVIS unseen test | RMSE: 25.95 -> 28.80 | Hurt |
| Use GeCo rather than DAVE in the zero-shot setup | FSCD147 zero-shot · FSCD147 test | MAE: 15.51 -> 13.30 | Helped |
Surprising findings
- Replacing the proposed loss with the standard unit-Gaussian loss raises validation MAE from 9.52 to 12.79 and RMSE from 43.00 to 59.33.
- The authors argue that high AP/AP50 can coexist with many false positives, so MAE/RMSE should be the primary counting measures.
Limitations
- Cannot process arbitrarily large images due to memory constraints, since it operates globally.
- Inference speed is not optimized; future work aims at a faster backbone.
- No error bars or statistical significance reported due to computational cost.
Experimental setup
Datasets: FSCD147 FSC147 FSCD-LVIS
Benchmarks: FSCD147 FSCD-LVIS FSCD147 multiclass subset
Metrics: MAE RMSE AP AP50
Baselines: LOCA CounTR SAFECount BMNet+ C-DETR SAM-C PSECO DAVE
Builds on: SAM DETR LOCA DAVE
Closest papers by meaning
- DAVE - A Detect-and-Verify Paradigm for Low-Shot Counting (CVPR 2024 · similarity 0.92)
- Point Segment and Count: A Generalized Framework for Object Counting (CVPR 2024 · similarity 0.91)
- Single Domain Generalization for Few-Shot Counting via Universal Representation Matching (CVPR 2025 · similarity 0.87)
- VLCounter: Text-Aware Visual Representation for Zero-Shot Object Counting (AAAI 2024 · similarity 0.86)
- T2ICount: Enhancing Cross-modal Understanding for Zero-Shot Counting (CVPR 2025 · similarity 0.86)
- Vision Transformer Off-the-Shelf: A Surprising Baseline for Few-Shot Class-Agnostic Counting (AAAI 2024 · similarity 0.86)
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.