Accurate keypoint localization of human pose needs diversified features: the high level for contextual dependencies and the low level for detailed refinement of joints. However, the importance of the two factors varies from case to case, but how to efficiently use the features is still an open problem. Existing methods have limitations in preserving low level features, adaptively adjusting the importance of different levels of features, and modeling the human perception process. This paper presents three novel techniques step by step to efficiently utilize different levels of features for human pose estimation. Firstly, an inception of inception (IOI) block is designed to emphasize the low level features. Secondly, an attention mechanism is proposed to adjust the importance of individual levels according to the context. Thirdly, a cascaded network is proposed to sequentially localize the joints to enforce message passing from joints of stand-alone parts like head and torso to remote joints like wrist or ankle. Experimental results demonstrate that the proposed method achieves the state-of-the-art performance on both MPII and LSP benchmarks.

Massive classification, a classification task defined over a vast number of classes (hundreds of thousands or even millions), has become an essential part of many real-world systems, such as face recognition. Existing methods, including the deep networks that achieved remarkable success in recent years, were mostly devised for problems with a moderate number of classes. They would meet with substantial difficulties, e.g., excessive memory demand and computational cost, when applied to massive problems. We present a new method to tackle this problem. This method can efficiently and accurately identify a small number of "active classes" for each mini-batch, based on a set of dynamic class hierarchies constructed on the fly. We also develop an adaptive allocation scheme thereon, which leads to a better tradeoff between performance and cost. On several large-scale benchmarks, our method significantly reduces the training cost and memory demand, while maintaining competitive performance.

Convolutional neural networks (CNNs) are usually built by stacking convolutional operations layer-by-layer. Although CNN has shown strong capability to extract semantics from raw pixels, its capacity to capture spatial relationships of pixels across rows and columns of an image is not fully explored. These relationships are important to learn semantic objects with strong shape priors but weak appearance coherences, such as traffic lanes, which are often occluded or not even painted on the road surface as shown in Fig. 1 (a). In this paper, we propose Spatial CNN (SCNN), which generalizes traditional deep layer-by-layer convolutions to slice-by-slice convolutions within feature maps, thus enabling message passings between pixels across rows and columns in a layer. Such SCNN is particular suitable for long continuous shape structure or large objects, with strong spatial relationship but less appearance clues, such as traffic lanes, poles, and wall. We apply SCNN on a newly released very challenging traffic lane detection dataset and Cityscapse dataset. The results show that SCNN could learn the spatial relationship for structure output and significantly improves the performance. We show that SCNN outperforms the recurrent neural network (RNN) based ReNet and MRF+CNN (MRFNet) in the lane detection dataset by 8.7% and 4.6% respectively. Moreover, our SCNN won the 1st place on the TuSimple Benchmark Lane Detection Challenge, with an accuracy of 96.53%.

Face grouping remains a challenging problem despite the remarkable capability of deep learning approaches in learning face representation. In particular, grouping results can still be egregious given profile faces and a large number of uninteresting faces and noisy detections. Often, a user needs to correct the erroneous grouping manually. In this study, we formulate a novel face grouping framework that learns clustering strategy from ground-truth simulated behavior. This is achieved through imitation learning (a.k.a apprenticeship learning or learning by watching) via inverse reinforcement learning (IRL). In contrast to existing clustering approaches that group instances by similarity, our framework makes sequential decision to dynamically decide when to merge two face instances/groups driven by short- and long-term rewards. Extensive experiments on three benchmark datasets show that our framework outperforms unsupervised and supervised baselines.

Speaker identification refers to the task of localizing the face of a person who has the same identity as the ongoing voice in a video. This task not only requires collective perception over both visual and auditory signals, the robustness to handle severe quality degradations and unconstrained content variations are also indispensable. In this paper, we describe a novel multimodal Long Short-Term Memory (LSTM) architecture which seamlessly unifies both visual and auditory modalities from the beginning of each sequence input. The key idea is to extend the conventional LSTM by not only sharing weights across time steps, but also sharing weights across modalities. We show that modeling the temporal dependency across face and voice can significantly improve the robustness to content quality degradations and variations. We also found that our multimodal LSTM is robustness to distractors, namely the non-speaking identities. We applied our multimodal LSTM to The Big Bang Theory dataset and showed that our system outperforms the state-of-the-art systems in speaker identification with lower false alarm rate and higher recognition accuracy.