CAVI-14: A Real-Time Vehicle Object Image Dataset for Autonomous Driving
Description
CAVI-14 (Camera-based Vehicle Image dataset with 14 categories) is a camera-based image dataset designed to support research in real-time object detection for intelligent transportation systems and autonomous driving. The dataset consists of 2,661 high-resolution images, collected using a Samsung Galaxy A15 smartphone camera around a university campus area in South Korea, capturing diverse urban traffic environments. The dataset includes 9,082 annotated objects across 14 vehicle-related object classes, specifically: • Vehicles & Transport: car, bus, pickup, lorry, ambulance, motorcycle, bicycle, e-bike • Pedestrian & Road Elements: pedestrian, zebra crossing, sidewalk, speed bump, traffic signal, road divider Each image is annotated with precise bounding boxes. The annotations are provided in two widely-used formats: • YOLO format (.txt) for real-time detection models • Pascal VOC format (.xml) for traditional detection pipelines The dataset is organized into three main directories: • train folder contains training images and their corresponding annotation files • test folder contains testing images and annotation files • annotations folder includes metadata in .xml format (object name, coordinates, image size, etc.) This structured and standardized format allows researchers to directly use the dataset for training and validating object detection models. A baseline evaluation using the YOLOv5s model achieved promising performance, confirming the dataset's suitability for benchmarking deep learning-based vehicle detection algorithms. CAVI-14 contributes a diverse and realistic set of labeled street-level traffic images, helping advance computer vision research in traffic monitoring, smart cities, and autonomous navigation.
Files
Steps to reproduce
Steps to reproduce: 1. Download the Dataset a. After accessing the dataset from Mendeley Data, extract the ZIP file. b. The dataset contains three main folders: train, test, and annotations. c. Each image file is in .jpg format, and labels are provided in both .txt (YOLO format) and .xml (Pascal VOC format). 2. Environment Setup a. Python ≥ 3.8 b. Install required libraries 3. Organize Dataset for Training your model 4. Prepare data.yaml File 5. Train your Model 6. Evaluation
Institutions
- Hankuk University of Foreign Studies - Global Campus
- Hankuk University of Foreign Studies