18 October 2022 in IEEE Transactions on Image Processing
Person re-identification is a problem of identifying individuals across non-overlapping cameras. Although remarkable progress has been made in the re-identification problem, it is still a challenging problem due to appearance variations of the same person as well as other people of similar appearance. Some prior works solved the issues by separating features of positive samples from features of negative ones. However, the performances of existing models considerably depend on the characteristics and statistics of the samples used for training. Thus, we propose a novel framework named sampling independent robust feature representation network (SirNet) that learns disentangled feature embedding from randomly chosen samples. A carefully designed sampling independent maximum discrepancy loss is introduced to model samples of the same person as a cluster. As a result, the proposed framework can generate additional hard negatives/positives using the learned features, which results in better discriminability from other identities. Extensive experimental results on large-scale benchmark datasets verify that the proposed model is more effective than prior state-of-the-art models.
8 March 2022 in Applied Intelligence
28 June 2021 in IEEE Access
We present a novel segmentation-based seam cutting algorithm to generate visually plausible high-resolution 360-degree video efficiently. While the demand for an efficient video stitching algorithm for generating immersive videos has increased, it has received limited attention in the literature. Furthermore, stitched videos often suffer from distorted objects, temporal inconsistency and time constraints. Thus, in this paper, we propose an efficient seam finding algorithm that preserves objects from distortion, minimizes temporal inconsistency, and reduces processing time. One of the fundamental steps in image and video stitching is the estimation of seam boundary. To do this, the proposed algorithm leverages a convolutional neural networks-based instance segmentation algorithm that provides more accurate object regions. It computes energy surfaces considering the regions and then estimates seam boundary by discovering a minimal energy path with minimal computations. We validate the proposed algorithm using real-world high-resolution 360-degree sequences. The experimental results verify that the proposed algorithm can produce seam boundaries that avoid objects with better temporary consistency. The proposed algorithm reduces the number of pixels passed through objects by approximately 30% on average compared to the existing algorithms. The qualitative comparisons furthermore demonstrate that the proposed algorithm consistently produces more perceptually pleasing results.