|
|
Vehicle Make and Model Recognition
Gregor, Adam ; Špaňhel, Jakub (oponent) ; Juránek, Roman (vedoucí práce)
In the practical part of the diploma thesis, the task of identifying the manufacturer and model of a vehicle (VMMR) was implemented. In the first part, a dataset of vehicles was compiled for machine learning purposes that consists of images from the Internet. This resulted in over 6 million images of cars, buses, motorbikes and trucks usable for the VMMR task. Next, as part of the experiments, a standard classification was used on a part of the dataset, when the encoder is followed by a classification layer implemented using a neural network. Also an approach with a supervised contrastive learning method, clustering embeddings from encoder for easier classification, was used. Since the first mentioned approach returned more accurate results, it was used in the further experiments. There, a larger portion of images from our dataset was used for training a classifier for the VMMR task. Other classifiers were trained on the Stanford Cars and Comprehensive cars datasets. Lastly, when comparing the functionality of the classifiers on different datasets we have found that the classifier trained on our dataset performed the best.
|
|
|
Vehicle Make and Model Recognition
Gregor, Adam ; Špaňhel, Jakub (oponent) ; Juránek, Roman (vedoucí práce)
In the practical part of the diploma thesis, the task of identifying the manufacturer and model of a vehicle (VMMR) was implemented. In the first part, a dataset of vehicles was compiled for machine learning purposes that consists of images from the Internet. This resulted in over 6 million images of cars, buses, motorbikes and trucks usable for the VMMR task. Next, as part of the experiments, a standard classification was used on a part of the dataset, when the encoder is followed by a classification layer implemented using a neural network. Also an approach with a supervised contrastive learning method, clustering embeddings from encoder for easier classification, was used. Since the first mentioned approach returned more accurate results, it was used in the further experiments. There, a larger portion of images from our dataset was used for training a classifier for the VMMR task. Other classifiers were trained on the Stanford Cars and Comprehensive cars datasets. Lastly, when comparing the functionality of the classifiers on different datasets we have found that the classifier trained on our dataset performed the best.
|
host ::
RSS.