SignBERT: Pre-Training of Hand-Model-Aware Representation for Sign Language Recognition

Abstract

Hand gesture serves as a crucial role during the expression of sign language. Current deep learning based methods for sign language understanding (SLU) are prone to over-fitting due to insufficient sign data resource and suffer limited interpretability. In this paper, we propose the first self-supervised pre-trainable SignBERT+ framework with model-aware hand prior incorporated. In our framework, the hand pose is regarded as a visual token, which is derived from an off-the-shelf detector. Each visual token is embedded with gesture state and spatial-temporal position encoding. To take full advantage of current sign data resource, we first perform self-supervised learning to model its statistics. To this end, we design multi-level masked modeling strategies (joint, frame and clip) to mimic common failure detection cases. Jointly with these masked modeling strategies, we incorporate model-aware hand prior to better capture hierarchical context over the sequence. After the pre-training, we carefully design simple yet effective prediction heads for downstream tasks. To validate the effectiveness of our framework, we perform extensive experiments on three main SLU tasks, involving isolated and continuous sign language recognition (SLR), and sign language translation (SLT). Experimental results demonstrate the effectiveness of our method, achieving new state-of-the-art performance with a notable gain.

BEST: BERT Pre-Training for Sign Language Recognition with Coupling Tokenization

“…As shown in Table 5, we compare with the previous methods. SignBERT (Hu et al 2021a) Our method outperforms SignBERT (Hu et al 2021a) with 4.89%, 5.96% and 9.28% Top-1 per-instance accuracy improvement on MSASL100, MSASL200 and MSASL1000, respectively. Notably, our method even achieves comparable performance with RGB-based methods.…”

“…Method MSASL100 MSASL200 MSASL1000 P-I P-C P-I P-C P-I (Yan, Xiong, and Lin 2018) 59 (Yan, Xiong, and Lin 2018) 90.0 SignBERT (Hu et al 2021a) 94.5 Ours 95.4…”

“…As shown in Table2, ST-GCN(Yan, Xiong, and Lin 2018) shows inferior performance compared with RGB-based methods, which may be attributed to the failure of pose detection. Compared with ST-GCN (Yan, Xiong, and Lin 2018), self-supervised learning methods, i.e., SignBERT(Hu et al 2021a) and Ours, relieve this issue by leveraging the stored statistics during pre-training.…”

Continuous Sign Language Recognition with Correlation Network

“…Sign language recognition methods can be roughly categorized into isolated sign language recognition [19,20,44] and continuous sign language recognition [5,7,34,35,38] (CSLR), and we focus on the latter in this paper. CSLR tries to translate image frames into corresponding glosses in a weakly-supervised way: only sentence-level label is provided.…”

Automatic dense annotation of large-vocabulary sign language videos

“…In the domain of continuous sign language recognition, in which the objective is to infer a sequence of sign glosses, prior work has explored HMMs [3,36] in combination with Dynamic Time Warping (DTW) [73], RNNs [18] and architectures capable of learning effectively from CTC losses [15,75]. Recently, sign representation learning methods inspired by BERT [20] have shown the potential to learn effective representations for both isolated [24] and continuous [76] recognition. Koller [35] provides an extensive survey of the sign recognition literature, highlighting the extremely limited supply of datasets with large-scale vocabularies suitable for continuous sign language recognition.…”