LVIS (Large Vocabulary Instance Segmentation) is a new dataset designed to enable the rigorous study of instance segmentation algorithms that can recognize a large vocabulary of object categories (>1000) and handle low-shot learning. The dataset aims to address the challenge of evaluating object detectors in scenarios with a large number of categories and limited training samples per category. LVIS will contain 164k images and ~2 million high-quality instance masks for over 1000 entry-level object categories. The dataset is structured as a federated dataset, where each category has a positive set and a negative set, allowing for efficient annotation and evaluation. The evaluation protocol is based on COCO-style instance segmentation and average precision (AP), ensuring continuity with existing benchmarks. The dataset construction process involves an iterative annotation pipeline that ensures high-quality segmentation masks and avoids ambiguous cases. Initial analysis shows that LVIS has a more diverse spatial distribution and smaller object sizes compared to COCO and ADE20K, highlighting its challenging low-shot nature. The dataset is available at <http://www.lvisdataset.org>, and the first LVIS Challenge will be held at the COCO Workshop at ICCV 2019.LVIS (Large Vocabulary Instance Segmentation) is a new dataset designed to enable the rigorous study of instance segmentation algorithms that can recognize a large vocabulary of object categories (>1000) and handle low-shot learning. The dataset aims to address the challenge of evaluating object detectors in scenarios with a large number of categories and limited training samples per category. LVIS will contain 164k images and ~2 million high-quality instance masks for over 1000 entry-level object categories. The dataset is structured as a federated dataset, where each category has a positive set and a negative set, allowing for efficient annotation and evaluation. The evaluation protocol is based on COCO-style instance segmentation and average precision (AP), ensuring continuity with existing benchmarks. The dataset construction process involves an iterative annotation pipeline that ensures high-quality segmentation masks and avoids ambiguous cases. Initial analysis shows that LVIS has a more diverse spatial distribution and smaller object sizes compared to COCO and ADE20K, highlighting its challenging low-shot nature. The dataset is available at <http://www.lvisdataset.org>, and the first LVIS Challenge will be held at the COCO Workshop at ICCV 2019.