Multi-Path Deep CNNs for Fine-Grained Car Recognition
Huibing Wang, Jinjia Peng, Yanzhu Zhao, Xianping Fu
Abstract
Along with the growing demands of intelligent traffic system, how to recognize the category information of a car from surveillance cameras has been an important task. Fine-grained car recognition is facing challenges mainly due to similar appearance of inter-class car images. Moreover, unlike other forms of object recognition, there are a large quantity of car models, that most other objects do not have, which causes fine-grained car recognition to be of vital importance but challenging. Recent research focuses on training a deep convolution neural network (DCNN) on a large dataset. Typically these methods largely rely on the holistic appearance of cars and may fail to train an optimal DCNN to comprehend information from multiple parts of a car. To address this issue, we carefully measure the effectiveness of different parts of a car and highlight the importance of car fronts for fine-grained car recognition. Then, considering multiple important parts from cars, we propose a novel multi-path DCNN model which is equipped with a 3-branch deep convolutional network to better exploit holistic as well as part information for fine-grained car recognition. Multiple parts from cars provide complementary information to boost performance. To further facilitate research on fine-grained car recognition, we also collect a large-scale dataset named “Multiple Parts from Cars” (MPF-Cars) which contains category annotation as well as car part information. We evaluate our proposed multi-path DCNN on MPF-Cars and another benchmark dataset CarFlag-563. Experiments demonstrate the effectiveness of our proposed method.