A Real-World Benchmark of Modern Object Detection Models
DetectionModelVersus is a newly-running benchmark project focused on training, evaluating, and comparing state-of-the-art object detection models using both quantitative metrics and real-world video inference.
Every week, two models are trained under identical conditions, evaluated, and compared. Results are published publicly and discussed on LinkedIn.
- YOLO11 (n, s, m, l, x)
- YOLO26 (n, s, m, l, x)
- RT-DETR (l, x)
- RF-DETR (Nano, Small, Medium, Large, XLarge, 2XLarge)
-
Week 1: YOLO11n vs RFDETRNano
-
Week 2: YOLO26n vs YOLO11s
-
Week 3: YOLO26s vs RFDETRSmall
-
Week 4: YOLO11m vs RFDETRMedium
-
Week 5: YOLO26m vs YOLO11l
-
Week 6: RFDETRLarge vs RTDETRl
-
Week 7: YOLO26l vs YOLO11x
-
Week 8: YOLO26x vs RTDETRx
-
Week 9: RFDETRXLarge vs RFDETR2XLarge
-
Week 10: Final recap & cross-model analysis
- mAP@50
- mAP@50:95
- Precision
- Recall
- Inference FPS
- Latency per frame
- GPU memory usage
- Stability across frames
- Performance on motion blur
- Small object detection
- Occlusion handling
- Two models trained per week
- Identical datasets, augmentations, and training schedules
- Side-by-side video inference
- Public metrics tables
- Results released every Thursday
All experiments are run on consistent hardware to ensure fair comparison. See hardware.md for full details.
- Weekly results are stored in the benchmarks directory
- Video inference outputs are stored in results/videos
Most benchmarks focus only on static accuracy metrics. DetectorModelVersus emphasizes:
- Practical deployment behavior
- Compute efficiency
- Visual detection quality in motion
This makes the results more useful for real-world production systems.
Agbaje Ayomipo
Data Scientist | Computer Vision Enthusiast
Weekly insights and breakdowns are shared on LinkedIn. Here