Vehicle Counting Model Report
V1 (Direct Regression) vs V2 (Density Map) | NVIDIA GB10 | 2026-04-12
V2 Mean MAE
1.523
vs V1: 1.744 (+12.7%)
V2 Mean RMSE
3.016
vs V1: 3.356 (+10.1%)
V2 Best Throughput
408 img/s
bs=4, bf16 | V1: 553 img/s
Validation Set
2,077
images | Train: 22,758 | 5 classes
GT vs Pred Scatter Plot (All Validation)
Each dot = 1 image. Closer to diagonal = more accurate. Hover for details.
Visual Comparison Gallery
Click any image to enlarge. Toggle between 4-panel composite and per-class breakdown.
4-Panel Composite
Per-Class Breakdown
4-Panel: Top-left: Original + GT points | Top-right: Error heatmap (red=missed, blue=over) | Bottom-left: Grid count (5x5) | Bottom-right: Density map (Inferno)
Per-Class: Each row = 1 class. Left column = GT density | Right column = Predicted density
Per-Class MAE Comparison
| Class | V1 MAE | V2 MAE | Delta | Change |
|---|---|---|---|---|
| car | 3.843 | 2.902 | -0.941 | +24.5% |
| motorcycle | 2.613 | 2.315 | -0.298 | +11.4% |
| truck | 0.775 | 0.795 | +0.021 | -2.7% |
| bus | 0.503 | 0.509 | +0.006 | -1.3% |
| bicycle | 0.987 | 1.095 | +0.108 | -10.9% |
| Mean | 1.744 | 1.523 | -0.221 | +12.7% |
Per-Class RMSE
| Class | V1 | V2 | Delta | Change |
|---|---|---|---|---|
| car | 6.650 | 5.262 | -1.388 | +20.9% |
| motorcycle | 5.546 | 4.840 | -0.706 | +12.7% |
| truck | 1.637 | 1.716 | +0.078 | -4.8% |
| bus | 0.958 | 1.044 | +0.085 | -8.9% |
| bicycle | 1.990 | 2.217 | +0.228 | -11.4% |
| Mean | 3.356 | 3.016 | -0.340 | +10.1% |
Car Error Distribution
| Percentile | V1 | V2 | Improvement |
|---|---|---|---|
| p50 | 1.95 | 1.66 | -14.9% |
| p75 | 4.84 | 3.75 | -22.5% |
| p90 | 9.66 | 7.22 | -25.3% |
| p95 | 13.63 | 9.79 | -28.2% |
| p99 | 23.50 | 16.67 | -29.1% |
Throughput Benchmark
Clean run on NVIDIA GB10 | Input: 384x384
V1 Peak
552.5 img/s
bs=4, bf16 | 1.81 ms/img | 117 MiB
V2 Peak
408.4 img/s
bs=4, bf16 | 2.45 ms/img | 149 MiB
| Batch | V1 img/s | V2 img/s | V2/V1 | V1 VRAM | V2 VRAM |
|---|---|---|---|---|---|
| 1 | 403.0 | 325.6 | 0.81x | 104 MiB | 106 MiB |
| 2 | 519.3 | 399.4 | 0.77x | 106 MiB | 108 MiB |
| 4 | 552.5 | 408.4 | 0.74x | 117 MiB | 149 MiB |
| 8 | 495.6 | 374.9 | 0.76x | 182 MiB | 247 MiB |
| 16 | 452.4 | 340.3 | 0.75x | 304 MiB | 443 MiB |
| 32 | 427.1 | 325.2 | 0.76x | 552 MiB | 830 MiB |
| 64 | 417.9 | 322.3 | 0.77x | 1047 MiB | 1607 MiB |
| 128 | 411.6 | 320.8 | 0.78x | 2038 MiB | 3159 MiB |
| 256 | 414.1 | 322.8 | 0.78x | 4017 MiB | 6264 MiB |
| 512 | 414.1 | 321.3 | 0.78x | 7976 MiB | 12472 MiB |
Conclusion
V2 Density Map is the recommended model
- Accuracy: Mean MAE +12.7%, car class +24.5%. Tail errors (p95, p99) improved ~28%
- Throughput: 408 img/s at bf16 — sufficient for 13+ simultaneous 30fps streams
- Spatial info: Density map enables ROI counting and error localization
- Trade-off: ~25% slower than V1; slightly worse on low-frequency classes
- Next steps: Class-weighted loss; higher resolution; fine-tune on domain-specific data
Generated 2026-04-12 | CVAT Project 24 (vehicle_seg) | gx10 NVIDIA GB10