Person Counting - Density Map Report
Model: ResNet50 FPN Density Map | CVAT Project 7 (person) | NVIDIA GB10 | 2026-04-13
Person MAE
1.798
Median error: 0.94
Person RMSE
3.218
Mean GT: 7.44 | Mean Pred: 7.15
Exact (err≤1)
52.1%
1,058 / 2,031 images
Validation Set
2,031
GT range: 1-92 persons
Training
50 epochs
Early stop patience=5 (not triggered)
Training Curve
GT vs Pred Scatter Plot
2,031 validation images. Closer to diagonal = more accurate.
Error Distribution
Error Percentiles
| Median (p50) | 0.94 | excellent |
| p75 | 2.22 | good |
| p90 | 4.39 | moderate |
| p95 | 6.70 | large |
| p99 | 13.86 | extreme |
Visual Comparison Gallery
Top-left: Original + GT points | Top-right: Error heatmap (red=missed, blue=over) | Bottom-left: Grid count 5x5 | Bottom-right: Density map. Click to enlarge.
Key Findings
Strong Performance
- MAE 1.798 on diverse validation set (hotels, airports, factories, outdoors)
- 52% exact (error ≤ 1 person), 82% within 3 persons
- Mean GT=7.44 vs Mean Pred=7.15 - no systematic bias (unlike vehicle cross-domain)
- GT range 1-92 persons handled well across densities
- Persistent overfitting warning suggests more augmentation or regularization could help
Generated 2026-04-13 | CVAT Project 7 (person) | Model: best_full.pt (epoch 45)