CarVideos: A Novel Dataset for Fine-Grained Car Classification in Videos

Abstract

We investigate the fine-grained object classification problem of determining make, model and year of car from a video. To this end we introduce a new dataset called CarVideos. It is a novel dataset for fine-grained object classification in videos. The CarVideos dataset contains over a million video frames annotated with bounding boxes around the visible cars as well as the specific year, make and model of each car. We implemented several state-of- the-art methods for object classification in videos and compared them using the dataset to establish a baseline performance level for future research. We also introduce a novel approach to fine-grained object classification in videos that combines a Single Shot Multibox Detector (SSD) with a single stream multi-region convolutional neural network (CNN). Our experiments show that our novel method significantly outperforms previous methods in terms of accuracy on the dataset. Our approach outperforms Temporal Segment Networks (TSN) and 3D Convolutional Networks, which are state-of-the-art on human action recognition in videos.