Multiple Instance Learning for Car Brand Classification in the Leboncoin Online Marketplace

This research project applies machine learning to a large-scale dataset of car images from the marketplace website Leboncoin. The project develops a model that classifies images from sales ads of cars with the brand of the car. The model combines the predictions of all the images within an ad to achieve an ad-wise classification. The motivation behind automatic image classification is to increase accuracy, identify mislabeled items and guide the user through the registration process of a sales ad. The methods developed in this project can be applied to a variety of visual classification tasks, such as other e-commerce sites, medical imagery and video classification. The project uses transfer learning to obtain faster training and higher accuracy due to a limited size of the dataset. The combination of predictions draws inspiration from multiple instance learning and shows that averaging of the predictions is a powerful bag-of-instance classifier. The classifier developed in this project gives an accuracy of 0.823 and 0.945 for image-wise and ad-wise respectively