Learning Metrics from Teachers: Compact Networks for Image Embedding
Authors: Lu Yu, Vacit Oguz Yazici, Xialei Liu, Joost van de Weijer, Yongmei Cheng, Arnau Ramisa
Accpted by Computer Vision and Pattern Recognition (CVPR), 2019
Abstract: Metric learning networks are used to compute image embeddings, which are widely used in many applications such as image retrieval and face recognition. For many real-world applications, the networks used to compute embeddings must be highly efficient, and therefore these applications cannot take advantage of the latest state-of-the-art deep networks. In this paper we study network distillation to efficiently compute image embeddings with small networks. Network distillation has been successfully used to improve image classification, but has hardly been explored for metric learning. To do so, we propose two new loss functions that model the communication of a deep teacher network to a small student network. We evaluate our system in several datasets, including CUB-200-2011 and Cars-196 and show that embeddings computed using small student networks perform significantly better than those computed using standard networks of similar size. Results on a very compact network (MobileNet-0.25), which can be used on mobile devices, show that the proposed method can greatly improve Recall@1 results from 27.5\% to 44.6\%. Furthermore, we investigate various aspects of distillation for embeddings, including hint and attention layers, semi-supervised learning and cross quality distillation.
Find out more
Weakly Supervised Domain-Specific Color Naming Based on Attention
(project webpage)
Authors: Lu Yu, Yongmei Cheng, Joost van de Weijer
Accpted by International Conference on Pattern Recognition (ICPR), 2018
Abstract: The majority of existing color naming methods focuses on the eleven basic color terms of the English language. However, in many applications, different sets of color names are used for the accurate description of objects. Labeling data to learn these domain-specific color names is an expensive and laborious task. Therefore, in this article we aim to learn color names from weakly labeled data. For this purpose, we add an attention branch to the color naming network. The attention branch is used to modulate the pixel-wise color naming predictions of the network. In experiments, we illustrate that the attention branch correctly identifies the relevant regions. Furthermore, we show that our method obtains state-of-the-art results for pixel-wise and image-wise classification on the EBAY dataset and is able to learn color names for various domains.
Find out more
Suitability of Real-Time Image under Complicated Environment Based on Contourlet in SMN
Authors: Lu Yu, Yongmei Cheng, Xialei Liu, Nan Liu
Accepted by Intelligent Systems and Knowledge Engineering (ISKE), 2015
Abstract: Judging whether the real-time image under complicated environment is suitable is a challenging problem in scene matching navigation, which contributes to ensure the navigation precision and decrease computational complexity. This paper proposes a novel method for analyzing the
suitability of real-time image under complicated environment based on Contourlet by taking advantage of the characteristic of multi-direction and multi-scale of Contourlet, where the complicated environment focus on motion blur, illumination variation, occlusion of cloud and fog. Firstly, real-time image is transformed on 4-layer Contourlet, and the obtained coefficients are parameterized by Generalized Gaussian Distribution, forming a 62 - dimension feature vector. Then the relationship between the feature vector and the objective evaluation index of suitability is trained by support vector
machine, to build the prediction model of suitability of realtime image under complicated environment. Finally, experiments are performed on image database picked from Google Earth. The experiments clearly demonstrate that the proposed algorithm is simple but effective for real-time image quality assessment in scene matching navigation.
Find out more
