In this paper, we dive into sign language recognition, focusing on the recognition of isolated signs. The task is defined as a classification problem, where a sequence of frames (i.e., images) is recognized as one of the given sign language glosses. We analyze two appearance-based approaches, I3D and TimeSformer, and one pose-based approach, SPOTER. The appearance-based approaches are trained on a few different data modalities, whereas the performance of SPOTER is evaluated on different types of preprocessing. All the methods are tested on two publicly available datasets: AUTSL and WLASL300. We experiment with ensemble techniques to achieve new state-of-the-art results of 73.84% accuracy on the WLASL300 dataset by using the CMA-ES optimization method to find the best ensemble weight parameters. Furthermore, we present an ensembling technique based on the Transformer model, which we call Neural Ensembler.

This work presents a novel transformer-based method for hand pose estimation—DePOTR. We test the DePOTR method on four benchmark datasets, where DePOTR outperforms other transformer-based methods while achieving results on par with other state-of-the-art methods. To further demonstrate the strength of DePOTR, we propose a novel multi-stage approach from full-scene depth image—MuTr. MuTr removes the necessity of having two different models in the hand pose estimation pipeline—one for hand localization and one for pose estimation—while maintaining promising results. To the best of our knowledge, this is the first successful attempt to use the same model architecture in standard and simultaneously in full-scene image setup while achieving competitive results in both of them. On the NYU dataset, DePOTR and MuTr reach precision equal to 7.85 mm and 8.71 mm, respectively.

In this article we tackle the problem of hand pose estimation when the hand is interacting with various objects from egocentric viewpoint. This entails a frequent occlusion of parts of the hand by the object and also self-occlusions of the hand. We use a Voxel-to-Voxel approach to obtain hypotheses of the hand joint locations, ensemble the hypotheses and use several post-processing strategies to improve on the results. We utilize models of prior hand pose in the form of Truncated Singular Value Decomposition (SVD) and the temporal context to produce refined hand joint locations. We present an ablation study of the methods to show the influence of individual features of the post-processing. With our method we were able to constitute state-of-the-art results on the HANDS19 Challenge: Task 2 - Depth-Based 3D Hand Pose Estimation while Interacting with Objects, with precision on unseen test data of 33.09 mm.