Vehicle keypoint detection and fine-grained classification using deep learning

Resumo

Vehicle keypoint detection and fine-grained classification systems have seen their capabilities evolve at an unprecedented rate, from poor performance to incredible results in a matter of a few years. The advent of convolutional neural networks and the availability of large amounts of data and progress in computational capabilities have allowed these and many other problems to be tackled and solved with very different approaches using increasingly complex models. This thesis focuses on the problems of keypoint detection and fine-grained classification of vehicles with a deep learning approach. After the analysis of the existing datasets to tackle both tasks, three new datasets have been built. The first one, oriented to the detection of keypoints in vehicles, is an improvement and extension of the famous PASCAL3D+ dataset, re-labelling part of it and adding new keypoints and images to provide more variability. The second is a vehicle make and model classification test set based on the PREVENTION dataset, a realworld driving scenario vehicle trajectory prediction dataset. Finally, a cross-dataset composed of common makes and models from three major vehicle classification databases, CompCars, VMMR-db and Frontal-103. The keypoint detection system is based on a human pose detection method that by using convolutional neural networks and deconvolutional layers generates, from an input image, a heat map for each keypoint. The network has been modified to fit the problem of keypoint detection in vehicles obtaining results that improve the state of the art without using complex architectures or methodologies. Additionally, the suitability of the PASCAL3D+ keypoints has been analysed, validating the proposal of new keypoints as a better alternative. The vehicle make and model classification system is based on the use of ImageNet pre-trained networks and fine-tuned for the vehicle classification problem. One of the problems detected in the state of the art is the saturation of the results in the existing datasets, which, moreover, are biased, limiting the generalisation capacity of the models trained with them. Multiple data learning and weighting techniques have been used to try to alleviate the impact of dataset bias. In order to assess the generalisation capabilities of the trained models in real situations, the PREVENTION test set has been used. Additionally, the cross-dataset has been used to evaluate the complexity of the existing datasets and the generalisation capabilities of the models trained with them. It is shown that competitive results can be achieved without the use of complex architectures and that a high quality dataset that adequately reflects the real world is needed in order to properly address the vehicle classification problem.