UA-DETRAC Vehicle Detection: Comparative Performance Report Using YOLOv8 and Faster R-CNN Author: Ömer Dursun Cengizhan Kınay Barış Başaran
- Introduction This report presents a comprehensive and detailed comparative analysis of object detection techniques applied to the UA-DETRAC dataset, focusing on two state-of-the-art methods:
- YOLOv8 (Nano and Small variants)
- Faster R-CNN with ResNet-50 Feature Pyramid Network (FPN) The objectives of this study were to: • Preprocess and structure the dataset for modern deep learning models • Train and evaluate multiple object detection architectures • Compare metrics including mAP, precision, recall, and inference speed • Iterate on training parameters and architecture variants to achieve optimal performance
- Dataset Description 2.1 Dataset Overview The UA-DETRAC dataset is a widely-used traffic surveillance benchmark composed of high-resolution video sequences. It includes detailed annotations for detecting and tracking multiple vehicle categories under various environmental conditions. For this project, video frames were extracted into still images and used with corresponding bounding box annotations. 2.2 Class Definitions The four categories annotated in this study are: • Car • Van • Bus • Others 2.3 Label Format Each image has an associated .txt annotation file in YOLO format: <class_id> <x_center> <y_center> All coordinates are normalized between 0 and 1. 2.4 Dataset Structure The following structure was used: YoloV8_detract/ ├── datasets/ │ ├── ua_detrac/ │ ├── images/ │ │ ├── train/ │ │ └── val/ │ ├── labels/ │ │ ├── train/ │ │ └── val/ │ └── data.yaml Each data.yaml file specified class names, training and validation image paths.
- Environment and Setup 3.1 Virtual Environment • OS: Windows 11 • Python version: 3.12.10 • Virtual Environment: python -m venv venv • Activated via: .\venv\Scripts\activate 3.2 Dependencies pip install ultralytics torch torchvision matplotlib 3.3 Hardware • GPU: NVIDIA GeForce RTX 3080 Ti
- YOLOv8 Implementation 4.1 YOLOv8n (Baseline) yolo task=detect mode=train model=yolov8n.pt data=datasets/ua_detrac/data.yaml epochs=50 imgsz=640 device=0 • mAP@0.5: 0.635 • mAP@0.5:0.95: 0.487 • Precision: 0.622 • Recall: 0.631 4.2 YOLOv8s (With Augmentation) yolo task=detect mode=train model=yolov8s.pt data=datasets/ua_detrac/data.yaml imgsz=960 epochs=30 patience=10 device=0 mosaic=1.0 hsv_h=0.015 hsv_s=0.7 hsv_v=0.4 • Early Stopping triggered at Epoch 19 • mAP@0.5: 0.697 • mAP@0.5:0.95: 0.541 • Precision: 0.678 • Recall: 0.666
4.3 Insights • Increasing image size (640 → 960) improved detection accuracy • Mosaic and HSV-based augmentation provided robustness to variance • YOLOv8s significantly outperformed YOLOv8n, especially on small and medium objects • Early stopping was used to terminate training at the optimal epoch and avoid overfitting
- Faster R-CNN Implementation 5.1 Architecture Used fasterrcnn_resnet50_fpn from torchvision.models.detection. Modified the classification head: in_features = model.roi_heads.box_predictor.cls_score.in_features model.roi_heads.box_predictor = FastRCNNPredictor(in_features, num_classes=4) 5.2 Dataset Loading Custom dataset loader was implemented to convert YOLO labels to absolute coordinates: xmin = (x_center - width/2) * image_width All bounding boxes and class labels were stored in dictionaries compatible with torchvision models. 5.3 Training Configuration Setting Value Epochs 50 Batch Size 4 Optimizer SGD Learning Rate 0.005 Device CUDA (RTX 3080 Ti)
5.4 Evaluation Results • mAP@0.5: 0.483 • mAP@0.5:0.95: 0.351 • Bus AP: 0.546
- Model Progression and Training Summary Model mAP@0.5 mAP@0.5:0.95 Precision Recall Notes Faster R-CNN 0.483 0.351 ~0.62 ~0.60 Trained with full dataset on RTX 3080 Ti YOLOv8n 0.635 0.487 0.622 0.631 Lightweight; fast but limited YOLOv8s (final) 0.697 0.541 0.678 0.666 Best performer with augmentations Notes: • The final YOLOv8s model with imgsz=960 and mosaic=1.0 delivered the best balance between precision and recall. • Faster R-CNN performance was robust but slightly lower due to slower convergence.
- Visual Results and Diagnostics • Plotted precision, recall, F1 and PR curves • Confusion matrices show class confusion; 'others' class benefited most from YOLOv8s • Labels correlogram indicated overrepresentation of 'van' class, slightly biasing predictions
- Conclusions The iterative refinement of object detection models on UA-DETRAC using YOLOv8 and Faster R-CNN revealed that: • YOLOv8s (small) is more effective than YOLOv8n or Faster R-CNN • Image resolution, augmentation strategies, and early stopping significantly influence training convergence and final accuracy • Faster R-CNN remains a solid anchor-based baseline trained with the full dataset on an RTX 3080 Ti GPU, but YOLOv8 showed superior inference speed and accuracy
- Future Work • Train Faster R-CNN with deeper backbones (e.g., ResNet-101) • Integrate ByteTrack or DeepSORT for real-time multi-object tracking • Enhance data augmentation with synthetic occlusion and blur • Quantize and export models for embedded deployment • Extend to nighttime or adverse weather condition detection tasks