Hierarchical Modified Fast R-CNN for Object Detection

Abstract

In object detection there is high degree of skewedness for objects' visual separability. It is difficult to distinguish object categories which demand dedicated classification. The deep convolutional neural networks (CNNs) are trained as N-way classifiers. As such considerable work is required towards leveraging hierarchical category structures. We present here Modified Fast region-based CNN (Mod Fast R-CNN) and Hierarchical Modified Fast region-based CNN (HMod Fast R-CNN) with deep CNNs being embedded considering categorical hierarchy. The easy classes are separated through coarse classifiers. The difficult classes are classified by fine classifiers. HMod Fast R-CNN is trained by initial components training which follows fine-tuning globally using multiple group discriminant analysis. The regularization is done using coarse category consistency. For large-scale recognition tasks, scalability is done considering conditional execution of fine category classifiers and layer parameters compression. Using MS-COCO (benchmark) CIFAR100 and VisualQA datasets we obtain good results. We build several different HMod Fast R-CNN versions where standard CNNs top-1 error is reduced significantly. HMod Fast R-CNN's performance superiority with other object detectors on PASCAL VOC 2007 and VOC 2012 datasets are also highlighted.