faezetta / vmmrdb

Vehicle Make and Model Recognition Dataset (VMMRdb)

This is an overview of the VMMR dataset introduced in "A Large and Diverse Dataset for Improved Vehicle Make and Model Recognition".

Overview

Despite the ongoing research and practical interests, car make and model analysis only attracts few attentions in the computer vision community. We believe the lack of high quality datasets greatly limits the exploration of the community in this domain. To this end, we collected and organized a large-scale and comprehensive image database called VMMRdb, where each image is labeled with the corresponding make, model and production year of the vehicle.

Description

The Vehicle Make and Model Recognition dataset (VMMRdb) is large in scale and diversity, containing 9,170 classes consisting of 291,752 images, covering models manufactured between 1950 and 2016. VMMRdb dataset contains images that were taken by different users, different imaging devices, and multiple view angles, ensuring a wide range of variations to account for various scenarios that could be encountered in a real-life scenario. The cars are not well aligned, and some images contain irrelevant background. The data covers vehicles from 712 areas covering all 412 sub-domains corresponding to US metro areas. Our dataset can be used as a baseline for training a robust model in several real-life scenarios for traffic surveillance.

_{The distribution of images in different classes of the dataset. Each circle is associated with a class, and its size represents the number of images in the class. The classes with labels are the ones including more than 100 images.}

Download

VMMRdb can be downloaded here.

Each image is labeled with the corresponding make, model and production year of the vehicle.

Some models referenced in our paper on VMMRdb-3036: Resnet-50 , VGG

Citation

If you use this dataset, please cite the following paper:

A Large and Diverse Dataset for Improved Vehicle Make and Model Recognition
F. Tafazzoli, K. Nishiyama and H. Frigui
In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops 2017. 

vmmrdb's Issues

Database Label

Finding error

Dear, faezetta. How are you?
I am gonna train vehicle make and model recognition application with deep learning. I tried with stanford dataset but i couldn't enough performance. I read the University of Louisville ' paper "Vehicle make and model recognition for intelligent
transportation monitoring and surveillance" and thought this dataset is very good for vmmr but i couldn't see anything in http://vmmrdb.cecsresearch.org/.
Would you let me know full url for downloading dataset?
Kind regards.
Thanks.

How to run the pre-trained models?

I want to run the pretrained ResNet .t7 model on a car image. How to run it? Do you have the inference code?

Download link do not work

As mentioned in title, download link do not work anymore. Is there possibility to fix that?

Model training code

Hi!

Awesome research. I was wondering if you could release the code for training the model?

Low validation accuracies when training on all data

Hello,
I hope that you are doing well.
I am trying to use the vmmrdb dataset to classify cars into their corresponding make, model, and year. I'm using the whole dataset (all 9170 classes) with a validation split of 0.2. However I am achieving very low validation accuracies (5-7%). I am using resnet50 with the pre-trained imagenet weights. I'm removing the layer and I'm adding a new dense layer with number of neurons = number of classes. I'm only training the last layer. I'm training for 200 epochs with a learning rate of 0.001.
I guess that I'm facing the multiplicity and ambiguity problems that are mentioned in the introduction of the paper so I was wondering if it's possible to have have access to the code that you used for training.This will help me in knowing the problem that i'm actually facing better.
Thanks in advance