Poribohon-BD: Bangladeshi local vehicle image dataset with annotation for classification

Vehicle Classification has become tremendously important due to various applications such as traffic video surveillance, accident avoidance, traffic congestion prevention, bringing intelligent transportation systems. This article presents ‘Poribohon-BD’ dataset for vehicle classification purposes in Bangladesh. The vehicle images are collected from two sources: i) smartphone camera, ii) social media. The dataset contains 9058 labeled and annotated images of 15 native Bangladeshi vehicles such as bus, motorbike, three-wheeler rickshaw, truck, wheelbarrow. Data augmentation techniques have been applied to keep the number of images comparable to each type of vehicle. For labeling the images, LabelImg tool by Tzuta Lin has been used. Human faces have also been blurred to maintain privacy and confidentiality. The dataset is compatible with various CNN architectures such as YOLO, VGG-16, R-CNN, DPM. It is available for research purposes at https://data.mendeley.com/datasets/pwyyg8zmk5/2.


Specifications
Computer Vision and Pattern Recognition Specific subject area Vehicle Classification Type of data 2D-RGB image (JPG) XML file How data were acquired Bangladeshi vehicle images are collected from two sources:

Value of the Data
• This dataset can be used to train deep learning models for vehicle detection, classification, and segmentation purposes. • Deep learning researchers interested in the area of vehicle identification, segmentation can be benefited using this dataset. More specifically, this dataset will benefit researchers in developing any traffic management applications for Bangladesh. • The dataset contains 9058 images of 15 Bangladeshi vehicles. It can be extended by increasing the number of images per class and adding some more types of vehicles. The extension of this dataset will improve and increase classification accuracy of deep learning models [1] . • This dataset can be used in multitudes of applications such as identifying unauthorized vehicles, detecting unfit vehicles, reducing exceeding speed, collecting highway toll, counting vehicles, receiving traffic information, checking empty spots in garages. • Identifying surrounding vehicles is important for a self-driving vehicle [2] . This dataset has no applicable limit in order to bring autonomous vehicle systems in Bangladesh. The advanced applications using this dataset might help the traffic police maintain traffic laws and make a more efficient traffic system.

Data Description
Poribohon-BD is an image dataset of 15 native vehicles of Bangladesh. The vehicles are: i) Bicycle, ii) Boat, iii) Bus, iv) Car, v) CNG, vi) Easy-bike, vii) Horse-cart, viii) Launch, ix) Leguna, x) Motorbike, xi) Rickshaw, xii) Tractor, xiii) Truck, xiv) Van, xv) Wheelbarrow. There are two types of data files in the dataset as follows: The data files are divided into 16 folders. Each folder contains images and annotation files of one single vehicle. The 'Multi-class Vehicles' folder contains images and annotation files of multiple types of vehicles. The number of images per class with other details is given in Table 2 .
There are multitudes of available datasets to train deep learning models such as COCO, Ima-geNet, MNIST, CIFAR10, PASCAL VOC. For vehicle detection and classification in developed countries, researchers have released several datasets such as KITTI dataset [3] , Waymo dataset [4] , Cityscapes dataset [5] , ApolloScape dataset [6] . A simple comparison of these public vehicle datasets with Poribohon-BD is given in Table 1 .

Data collection
To develop any traffic management application for developing countries like Bangladesh, researchers will need a vast amount of images of different native vehicles. Thus, presenting Poribohon-BD dataset in this article aims to provide such a collection. The images are collected from two different sources: 1) Smartphone Cameras: The images of this dataset have been captured using smartphone cameras from different locations, roads, highways, beaches of Bangladesh. Both images and videos were captured using smartphone cameras. Selected frames from the video files are then converted in still images. Different views, backgrounds, weather conditions, scenarios have been considered while taking the pictures to increase variance in the data. 2) Social Media: Around 40 0 0 images are collected from social media (facebook). The images are taken from different facebook profiles with personal consents. Moreover, privacy issues are resolved by hazing the faces and any personal information.

Data pre-processing
After the data collection phase, all of the images have been converted in JPG format. Due to maintaining privacy and confidentiality, human faces or any other kind of personal information have been blurred in the images.

Data augmentation
Data Augmentation is a popular process in machine learning for increasing the amount and diversity of data. It is a popular solution to reduce overfitting in small datasets [7] . In Poribohon-BD dataset, few data augmentation techniques such as flipping, cropping, color space transformation have been applied to generate 1791 new images. The augmented images are also in JPG format.

Data annotation
An annotation file represents the location of an object in an image by containing the coordinates and label of that object [8] . In this last phase, popular annotation tool LabelImg by Tzuta Lin has been used to cautiously label the images. First of all, each image is opened in this tool one by one. Then, a rectangular shape has been drawn manually to the boundary of an object to specify its exact location in that image by X-Y coordinates. Finally, a label has been assigned such as bus, truck, bicycle to each object. In LabelImg, annotated values are saved as XML files in PASCAL VOC format [9] .

Ethics Statement
The reuse of images from Facebook complies to the terms of use. All the images were acquired with the consent of the people, groups or organizations.