Weakly-supervised vehicle detection and classification by convolutional neural network
Changyu Jiang, Bailing Zhang · 2016
Vehicle detection and vehicle type/make classification have been attracting more research in recent years. Previous methods for vehicle detection typically rely on large number of annotated training images by object bounding boxes, which is expensive and often subjective. In this paper, we propose a vehicle detection and recognition system by applying weakly-supervised convolutional neural network (CNN), with training relying only on image-level labels. Experiments were conducted on a datasets acquired from field-captured traffic surveillance cameras, with vehicle classification performance mAP 98.79% and accuracy 98.28%, and vehicle detection performance mAP 85.26%.