Input Data Partitioning For Detection Acceleration And Siamese Learning

AlexNet achieved the best results in the ImageNet 2012 competition,outperform-ing all other approaches by a large margin,which then marked the opening of an era named deep learning.Nowadays,methods based on deep learning have achieved great success in various areas.In computer vision,as a representative of deep learning,Con-volutional Neural Network(CNN)have achieved the best performance in nearly all topics,such as image classification,object detection and image segmentation.How-ever,as the traditional CNNs are complex,it is difficult to implement them efficiently or change their basic architectures,both of which would become a barrier for the popu-larity of deep models.In this paper,we propose to exploit the partition of input images,so that we could utilize the properties of different neural networks,run them more ef-ficiently and improve their performance.Our works can be summarized as follows:We propose Adaptive Feeding(AF)to combine a fast(but less accurate)detector and an accurate(but slow)detector,by adaptively determining whether an image is easy or hard and choosing an appropriate detector for it.Experiments show that our approach leads to 50%speed up while maintaining the same accuracy.Since CNNs have achieved amazing results on existing tasks,we try to explore and solve another harder problem:subtle attribute recognition.We first evaluate classical models on this question and then propose a new pairwise learning strategy to learn better representations.Experiments show that our approach surpasses baseline methods by a large margin and achieves better results when compared with human-beings.