Wei Hao

The 2D human pose estimation plays an important role in human‐computer interaction and action recognition. Although the method based on high‐resolution network has superior performance, there is still room for improvement in terms of speed and lightweight. Here, a LiteDEKR, a 2D pose estimation method that combines lightweight and accuracy, is proposed by designing a lightweight network based on DEKR and constructing two scientifically valid loss functions. The method, constructs a multi‐instance bias regression loss that matches the true distribution of keypoint bias, improves the accuracy of bias regression, and constructs a keypoint similarity loss with the object keypoint similarity index of keypoints as the optimization objective to achieve end‐to‐end training of the network. In addition, this paper has designed a lightweight DEKR, using LitePose as the backbone network. With the optimization of the above two loss functions, LiteDEKR not only achieves lightweight but also has high accuracy. Comparative experiments on the COCO and CrowdPose datasets show that compared to the current state‐of‐the‐art Contextual Instance Decoupling, LiteDEKR achieves a similar accuracy with only 10% of its network complexity. It also shows better robustness to low‐resolution input images.

Video object segmentation has been extensively investigated in computer vision recently because of its wide range of applications. The key factor of the segmentation is the construction of the spatiotemporal coherence. Inaccurate motion approximation as a measurement of the coherence usually leads to an inaccurate segmentation result. To obtain an accurate segmentation result, a low‐rank sparse (LRS)‐based approach is proposed. Regarding each superpixel as an element, this algorithm has a good segmentation accuracy compared with other pixel‐level algorithms. Each element can be represented by the sparse linear combinations of dictionary templates, and this algorithm capitalises on the inherent low‐rank structure of representations that are learnt jointly. The represented coefficients construct an affinity matrix which measures the elements’ similarity between the current frame and the templates in the dictionary. For video object segmentation, a principled spatiotemporal objective function that uses LRS saliency term to propagate information between frames. Furthermore, an online parameter updating scheme is proposed to enhance the system's robustness. The online model propagates information forward without the need to access future frames. Evaluations on many challenging sequences demonstrate that the authors' approach outperforms the state‐of‐the‐art methods in terms of object segmentation accuracy.

Graphcut based algorithm is adopted in many video object segmentation systems because different terms can be probabilistically fused together in a framework. Constructing spatio-temporal coherences is an important stage in segmentation systems. However, many steps are involved when computing a key term with good discriminative power. If the cascade steps are adopted, the inaccurate output of the previous step will definitely affect the next step, leading to inaccurate segmentation. In this paper, a key term which is computed by a single framework referred as Boundary-Constrained Low-Rank Sparse Representation (BCLRSR) is proposed to achieve the accurate segmentation. By treating the elements as linear combinations of dictionary templates, low-rank sparse optimization is adopted to achieve the spatiotemporal saliency. For adding the spatial information to the low-rank sparse model, a boundary constraint is adopted in the framework as a Laplacian regularization. A BCLRSR saliency is then obtained by the represented coefficients which measure the similarity between the elements in the current frame and the ones in the dictionary. At last, the object is segmented by minimizing the energy function which is formalized by the spatio-temporal coherences. Experiments on some public datasets show that our proposed algorithm outperforms the state-of-the-art methods.