SignNet-1M: Large-Scale Multilingual Sign Language Video Dataset with Downstream Benchmarks

A multilingual sign language video dataset called SignNet-1M has been released, spanning American, Chinese, and German sign languages. The dataset uses synthetic augmentations to improve model robustness against real-world shifts in viewpoint, background, and signer identity. The dataset was introduced in a paper submitted on 23 June 2026 to the arXiv preprint server [1]. Sign language models are typically trained on data captured under constrained conditions, which limits their ability to handle the diversity of real-world recordings [2]. SignNet-1M addresses this by synthesizing realistic variations along three axes [2]. The first is novel-view rendering, which uses 3D Gaussian Splatting to simulate rotation and zoom [2]. The second is scene and identity editing, where diffusion models replace backgrounds and substitute signers while preserving the original sign motion and linguistic content [2]. The third axis consists of post-rendering augmentations that emulate capture and compression artifacts, including pose and temporal perturbations and video-level corruptions [2]. The dataset covers American Sign Language, Chinese Sign Language, and German Sign Language (DGS) [2]. The researchers also provide a unified benchmark suite for downstream tasks such as translation and recognition, along with ablation studies that isolate the contribution of each augmentation component [2]. Experiments across multiple model backbones showed that training with SignNet-1M consistently improved generalization under cross-view, cross-background, cross-identity, and post-rendering shifts, while maintaining strong in-distribution performance [2]. The full dataset, augmentation pipeline, and benchmark are publicly available at the project's website [2]. The work reflects a broader trend in machine learning where augmented or synthetic data is used to supplement real-world datasets, a strategy that has been explored in other domains such as catalyst informatics to improve model transferability [4]. The release of SignNet-1M provides a resource for developing sign language recognition and translation systems that are more robust to the varied conditions encountered outside the laboratory.