weiguangzhao/synthverse

SynthVerse is a research dataset built for a computer vision task called point tracking, which means following the same physical point across many video frames as a camera or scene moves. The dataset comes from a paper accepted at SIGGRAPH 2026, a major graphics research conference, and it was created synthetically, meaning the scenes and camera motion were generated by computer rather than filmed in the real world. This lets the creators produce a large and varied set of examples for testing how well tracking algorithms work under different conditions. The repository itself is mostly a pointer to two hosted collections on Hugging Face, a platform for sharing datasets and models: the SynthVerse Benchmark, used to evaluate tracking systems, and the full SynthVerse Dataset, used for training them. Each sequence in the dataset includes RGB color images, matching depth images that record distance from the camera, camera position and intrinsic information, and the tracked 2D and 3D coordinates of points over time. A dataloader, the code that reads this data into a training pipeline, is included, though the code used to generate the synthetic scenes themselves is not yet released. This project is aimed at researchers and engineers working on computer vision, robotics, or 3D scene understanding who need training or evaluation data for point tracking models. It is not a general purpose tool or application. There is no installation guide beyond downloading the dataset files and using the provided dataloader, and no license information is stated in the README. The authors credit several other open source point tracking projects that influenced this work, including Kubric, PointOdyssey, TAPNET, and TAPIP3D. Anyone using the dataset in their own research is asked to cite the associated paper.