Vehicle-10 dataset: Vehicle classification

Dataset Overview

The Vehicle-10 dataset foucuses on vehicle classification task. Specifically, we collected 30006 vehicle images from internet and divided them into 10 categories, e.g., bicycle, boat, bus, car, helicopter, minibus, motorcycle, minibus, taxi, train and truck. These images vary in resolution, ranging from 47pix × 36pix to 4101pix × 2651pix. In the following, we give an example of different vehicle's images.

Features

The size of training set includes 28804 images and the validation set consists of 7202 images. The label id and number of samples for each category are listed below:

Category	bicycle	boat	bus	car	helicopter	minibus	motorcycle	taxi	train	truck
Label Id	0	1	2	3	4	5	6	7	8	9
Size	1618	8897	4064	8540	668	1477	4438	908	1682	3714

Download

We also provide two data compression files for you to choose from:

• Linux: vehicle-10.tar.gz

• Windows: vehicle-10.zip

After downloading and unzipping the above files, you can get the following directory structure. Concretely, The files named train_meta.json and valid_meta.json record the metadata of training set and validation set.

.
├── bicycle
├── boat
├── bus
├── car
├── helicopter
├── minibus
├── motorcycle
├── taxi
├── train
├── truck
├── train_meta.json
└── valid_meta.json

Evaluation of Machine Learning

The evaluation code for the Vehicle-10 can be found in the ./src directory. We briefly evaluated LeNet, ResNet (e.g., ResNet9, ResNet-18, ResNet-34, ResNet-50) and VGG (e.g., VGG-16, VGG-19) model architectures on Vehicle-10 dataset. Futhermore, we report the experimental results as following:

Model	LetNet5	ResNet9	ResNet18	ResNet34	ResNet50	VGG16	VGG19
Training Loss	1.340	0.212	0.385	0.349	0.624	0.560	0.592
Testing Loss	1.492	1.093	1.221	1.297	1.556	1.081	1.281
Accuracy	51.125%	70.217%	72.799%	74.660%	67.509%	78.645%	77.951%

Traing parameters are epoch=50, batchsize=128 and img_size=224(for LeNet5 and ResNet9, img_size=32). Specifically, you can obtain the results by executing the scripts.

cd scripts/

sh lenet5.sh
sh resnet9.sh
sh resnet18.sh
sh resnet34.sh
sh resnet50.sh
sh vgg16.sh
sh vgg19.sh

Evaluation of Federated Learning

We also evaluated some popular federated algorithms (e.g., FedAvg, FedProx, FedNova) on Vehicle-10 dataset in Non-IID setting. Assumed that 100 clients participate in federation learning initiated by the server. In each communication round, the server selects 20% of the clients to participate in training process. We have the following results.

Algorithm	FedAvg	FedProx	FedNova
Final Accuracy(ρ=20%)	62.18 ± 4.03	56.03 ± 6.73	48.69 ± 8.47
Final Accuracy(ρ=30%)	69.33 ± 6.04	69.67 ± 6.83	68.65 ± 5.08

Note: ρ=20% represents that each client will be randomly assigned 20% of number of categories. We report the mean and standard deviation for the final accuracy over 3 runs.

You can run the scripts for the quickstart.

cd scripts/

sh fedavg.sh
sh fedprox.sh
sh fednova.sh

Citation

If you find Vehicle-10 dataset to be useful in your own research, please consider citing the following paper:

@article{zhai2024fedrav,
  title={FedRAV: Hierarchically Federated Region-Learning for Traffic Object Classification of Autonomous Vehicles},
  author={Zhai, Yijun and Zhou, Pengzhan and He, Yuepeng and Qu, Fang and Qin, Zhida and Jiao, Xianlong and Liu, Guiyan and Guo, Songtao},
  journal={arXiv preprint arXiv:2411.13979},
  year={2024}
}

Acknowledgments

The Vehicle-10 dataset could not be collected without the following contributors. But we also do a lot of work on aligning and filtering for these scattered data.

• 5 vehichles for classification

• Vehicle Classification

• TAU Vehicle Type Recognition Competition