Citation Context

... of human-body part localization has made significant progress in recent years. This has been in part due to the success of DeepLearning architectures - specifically Convolutional Networks (ConvNets) =-=[21, 14, 22, 5]-=- - but also due to the availability of ever larger and more comprehensive datasets [1, 16, 20] (our model’s predictions for difficult examples from [1] are shown in Figure 1). A common characteristic ...

Citation Context

...ed and not learnt. Convolutional networks (ConvNets) are by far the best performing algorithms for many vision tasks. The state-ofthe-art methods for human-pose estimation are also based on ConvNets (=-=[40, 22, 39, 23, 11]-=-). Toshev et al. [40] formulate the problem as a direct regression to joint location. Chen et al. [11] improve over [40] by adding an image dependent spatial prior. Jain et al. [22] train an image pat...

Citation Context

...d in Section 3; Section 5.2 evaluates our approach on the regularization path estimation. Datasets. We evaluate our methods on four datasets for different structured prediction tasks: OCR (Taskar et al., 2003) for handwritten character recognition, CoNLL (Tjong Kim Sang & Buchholz, 2000) for text chunking, HorseSeg (Kolesnikov et al., 2014) for binary image segmentation and LSP (Johnson & Everingham, 2010) for pose estimation. The models for OCR and CoNLL were provided by Lacoste-Julien et al. (2013). We build our model based on the one by Kolesnikov et al. (2014) for HorseSeg, and the one by Chen & Yuille (2014) for LSP. For OCR and CoNLL, the max oracle consists of the Viterbi algorithm (Viterbi, 1967); for HorseSeg – in graph cut (Boykov & Kolmogorov, 2004), for LSP – in belief propagation on a tree with messages passed by a generalized distance transform (Felzenszwalb & Huttenlocher, 2005). Note that the oracles of HorseSeg and LSP require positivity constraints on a subset of the weights in order to be tractable. The BCFW algorithm with positivity constraints is derived in App. H. We provide a detailed description of the datasets in App. I with a summary in Table 1. The problems included in our e...

Citation Context

...mes of temporal aggregation. The extraction of proposed Posebased Convolutional Neural Network (P-CNN) features is illustrated in Figure 1. Pose estimation in natural images is still a difficult task =-=[7, 37, 42]-=-. In this paper we investigate P-CNN features both for automatically estimated as well as manually annotated human poses. We report experimental results for two challenging datasets: JHMDB [19], a sub...

Citation Context

...ined categorisation techniques rely on parts, it is important to look at pose estimation. To benchmark human pose estimation we used the state-of-theart, articulated pose estimator of Chen and Yuille =-=[8]-=-. This method extends Yang and Ramanan’s work [41] to use deep features. The code is publicly available. Two experiments were run using a single train/test split. In the first experiment, the pose mod...

Citation Context

...volution between features maps in the same layer, and it is also integrated with feedforward propagation in neural networks. Finally, a neural network implementation of endto-end learning CRF-CNN is provided. Its effectiveness is demonstrated through experiments on two benchmark datasets. 1 Introduction A lot of efforts have been devoted to structure design of convolutional neural network (CNN). They can be divided into two groups. One is to achieve higher expressive power by making CNN deeper [19, 10, 20]. The other is to model structures among features and outputs, either as post processing [6, 2] or as extra information to guide the learning of CNN [29, 22, 24]. They are complementary. Human pose estimation is to estimate body joint locations from 2D images, which could be applied to assist other tasks such as [4, 14, 26] The very first attempt adopting CNN for human pose estimation is DeepPose [23]. It used CNN to regress joint locations repeatedly without directly modeling the output structure. However, the prediction of body joint locations relies both on their own appearance scores and the prediction of other joints. Hence, the output space for human pose estimation is structured....

Citation Context

...ion. While in the experiments on 2D, 3D pose and viewpoint estimation, we construct our model based on the Mixture-of-Parts structure as [41] of 10 part types, based on HOG feature and CNN feature as =-=[7]-=-. We use the Caffe [23] to compute the CNN feature. When learning the scene-domain geometric subspace B, we follow the NRSFM techniques [9] to set the geometric subspace bases number K = 5 for object ...

Citation Context

...ity to learn hierarchical abstractions of data (Zeiler & Fergus, 2014). While this invariance is clearly desirable for high-level vision tasks, it can hamper low-level tasks, such as pose estimation (=-=Chen & Yuille, 2014-=-; Tompson et al., 2014) and semantic segmentation - where we want precise localization, rather than abstraction of spatial details. There are two technical hurdles in the application of DCNNs to image...

Citation Context

... spatial interactions among body parts. A novel, non-linear joint regressor model was proposed in [6], which handles typical ambiguities of tree based models quite well. More recently, deep learning (=-=[36, 35, 18, 4]-=-) has also been introduced for human pose estimation. For video based human pose estimation in unconstrained scenes, some early research adopted the tracking-by-detection framework ([1, 17, 25]). More...

Citation Context

...bear? Without an accurate understanding of the interaction, we will not be able to generate informative image descriptions besides a bag of objects. Despite significant advances in recognizing humans =-=[2]-=- and objects [14], the state of the art of HOI recognition in images is still far from the demands of real-world applications. A key bottleneck is the limited number of HOI cate1In this paper, we will...