LVIS is a new dataset for Large Vocabulary Instance Segmentation, containing over 1000 object categories and 164k images with ~2 million high-quality instance segmentation masks. The dataset is designed to address the challenge of instance segmentation in the low-sample regime, where many categories have few training examples. LVIS is built using a federated dataset approach, where each category is represented in separate subsets of the dataset, allowing for efficient annotation and evaluation. The dataset includes a wide range of object categories, with a long tail of rare categories, making it a valuable resource for evaluating instance segmentation algorithms. The dataset is available at http://www.lvisdataset.org. LVIS is designed to enable research on large vocabulary instance segmentation, with a focus on methods that can handle the open problem of low-shot learning. The dataset includes a detailed evaluation protocol, with metrics such as mask average precision (AP) and boundary quality. The dataset is also used to evaluate the performance of existing instance segmentation algorithms, including Mask R-CNN. The results show that LVIS provides a challenging benchmark for instance segmentation, with a significant portion of categories having few training examples. The dataset is expected to be used in future research and benchmarking efforts, helping to advance the field of instance segmentation and low-shot learning.LVIS is a new dataset for Large Vocabulary Instance Segmentation, containing over 1000 object categories and 164k images with ~2 million high-quality instance segmentation masks. The dataset is designed to address the challenge of instance segmentation in the low-sample regime, where many categories have few training examples. LVIS is built using a federated dataset approach, where each category is represented in separate subsets of the dataset, allowing for efficient annotation and evaluation. The dataset includes a wide range of object categories, with a long tail of rare categories, making it a valuable resource for evaluating instance segmentation algorithms. The dataset is available at http://www.lvisdataset.org. LVIS is designed to enable research on large vocabulary instance segmentation, with a focus on methods that can handle the open problem of low-shot learning. The dataset includes a detailed evaluation protocol, with metrics such as mask average precision (AP) and boundary quality. The dataset is also used to evaluate the performance of existing instance segmentation algorithms, including Mask R-CNN. The results show that LVIS provides a challenging benchmark for instance segmentation, with a significant portion of categories having few training examples. The dataset is expected to be used in future research and benchmarking efforts, helping to advance the field of instance segmentation and low-shot learning.