The vehicle orientation dataset contains more than one million annotations of vehicles with orientation in more than 200,000 images. It reduces the need of a secondary neural network to classify orientation by simultaneously providing both vehicle class and direction. Here are our published papers at ISPRS Journal of Photogrammetry and Remote Sensing and IEEE International Conference on Big Data 2021:
Real-time citywide reconstruction of traffic flow from moving cameras on lightweight edge devices»
Citywide reconstruction of cross-sectional traffic flow from moving camera videos»
Download Vehicle Orientation Dataset
Download Experiment Dataset (Video and GPS)
We'll continue adding various object detection models trained on the vehicle orientation dataset and the synthetic vehicle orientation dataset. Open an issue if you need some specific pre-trained weights.
| Framework/Network | Size | Dataset | Model | Download weights |
|---|---|---|---|---|
| YOLOv4 (darknet) | 608x608 | Vehicle Orientation Dataset | YOLOv4 | YOLOv4 weights |
| YOLOv5 (Ultralytics) | 640x640 | Vehicle Orientation Dataset | YOLOv5l | YOLOv5l weights |
| YOLOv5 (Ultralytics) | 640x640 | Vehicle Orientation Dataset | YOLOv5x | YOLOv5x weights |
| YOLOv5 (Ultralytics) | 1280x1280 | Vehicle Orientation Dataset | YOLOv5x6 | YOLOv5x6 weights |
| YOLOv5 (Ultralytics) | 1280x1280 | Vehicle Orientation Dataset | YOLOv5l6 | YOLOv5l6 weights |
All vehicles in the vehicle orientation dataset are labeled with both vehicle class (five categories) and its orientation (three types).
The five classes of vehicles are:
- Car
- Bus
- Truck
- Motorcycle
- Bicycle
The three types of orientations are:
- Front
- Back
- Side
So the vehicle orientation dataset has a total of 15 classes of vehicles with orientation such car_back, car_front, car_side, bus_back, bus_front, etc.
Annotations per class in the vehicle orientation dataset follows the long-tail distribution as commonly seen in other vehicle detection data sets.
The vehicle orientation dataset is hosted on AWS S3 (Asia-pacific, Tokyo) bucket. Since the overall size of the dataset is quite big (~100GB), we have split the vehicle orientation dataset into five parts for convenience of users. Part 1 to Part 4 together contain 200,000 images (50,000 x 4) and Part 5 has 13,714 images.
Please note that the annotations are provided in YOLO format style (darknet). There is a .txt-file for each .jpg-image-file - in the same directory and with the same name. Each line contains the class and bounding box coordinates for a vehicle in the image. If there are multiple vehicles in the image, the number of lines will increase accordingly.
<object-class> <x_center> <y_center> <width> <height>
where:
<object-class>- integer object number from0to(classes-1). Mapping file can be downloaded from here: Vehicle Orientation Classes<x_center> <y_center> <width> <height>- float values relative to width and height of image, it can be equal from(0.0 to 1.0]- For example:
<x> = <absolute_x> / <image_width>or<height> = <absolute_height> / <image_height> - Attention:
<x_center> <y_center>- are center of rectangle (Not top-left corner)
For example, for SUG007M5MX5JAZGUI4EI.jpg in vehicle-orientation-5 we have the corresponding annotation file SUG007M5MX5JAZGUI4EI.txt containing:
2 0.650000 0.573148 0.018750 0.027778
6 0.864062 0.449537 0.265625 0.793519
1 0.300000 0.581481 0.068750 0.051852
0 0.558594 0.625463 0.110937 0.217593
The first column represents the class such as car_front, car_back, etc. 2 in the first row means car_front, 6 in the second row is truck_back, and so on. Please check Vehicle Orientation Classes file for all 15 classes.
- Part-1, 50,000 images
- Part-2, 50,000 images
- Part-3, 50,000 images
- Part-4, 50,000 images
- Part-5, 13,714 images
Distributed under the MIT License. See LICENSE.txt for more information.
For any question and support, please create an issue on GitHub or write to the author here:
Ashutosh Kumar - ashutosh[at]iis.u-tokyo.ac.jp
@article{kumar2022real,
title={Real-time citywide reconstruction of traffic flow from moving cameras on lightweight edge devices},
author={Kumar, Ashutosh and Kashiyama, Takehiro and Maeda, Hiroya and Omata, Hiroshi and Sekimoto, Yoshihide},
journal={ISPRS Journal of Photogrammetry and Remote Sensing},
volume={192},
pages={115--129},
year={2022},
publisher={Elsevier}
}
@inproceedings{kumar2021citywide,
title={Citywide reconstruction of cross-sectional traffic flow from moving camera videos},
author={Kumar, Ashutosh and Kashiyama, Takehiro and Maeda, Hiroya and Sekimoto, Yoshihide},
booktitle={2021 IEEE International Conference on Big Data (Big Data)},
pages={1670--1678},
year={2021},
organization={IEEE}
}
We are organizing IEEE BigData Cup Challenge on "Vehicle class and orientation detection in the real-world using synthetic images from driving simulators." Participate to win cash prizes and free registration to attend the IEEE BigData Cup 2022 conference this year in Osaka, Japan.
- Training dataset (train-1): train-1
- Training dataset (train-2): train-2
- Test dataset (test-1): test-1
- Test dataset (test-1): test-2
The distribution of annotations in the training dataset (train-1 and train-2) is as shown below:
| Class | Number of annotations |
|---|---|
| car_front | 42273 |
| car_back | 35017 |
| car_side | 13131 |
| truck_front | 1995 |
| truck_back | 2667 |
| truck_side | 1220 |
| motorcycle_front | 770 |
| motorcycle_back | 1476 |
| motorcycle_side | 2614 |
| cycle_front | 498 |
| cycle_back | 1284 |
| cycle_side | 1881 |