Since its launch, the OpenDataLab platform has covered more than 4,500 datasets and over 1,200 task types, attracting widespread attention from friends in the industry.
This article summarizes the ultra-large-scale data sets in multiple fields on the OpenDataLab platform, including dynamic 4D face data sets, high-fidelity synthetic human body data sets, multi-modal human body data sets, High-quality object scene data, large-scale game-generated datasets, brand-new datasets for testing the generation of new human perspectives, the first real-world and largest 3D lane datasets, etc., for everyone to enjoy.
No.1 OpenXD-RenFace
● Domain: dynamic 4D face dataset
● Introduction:
RenFace is the only large-scale full-head human head dataset containing rich fine-grained hairstyles, and also contains phoneme-balanced speech videos. The number of IDs reaches 500+. The total number of acquisition frames reaches 80M, the number of viewing angles reaches 60, and the image resolution reaches 2k. The data set contains multiple races, multiple age groups, and multiple nationalities to be collected, with multiple high-fine-grained hairstyles and hair colors, and audio and video collection of the reading corpus at the same time. The dynamic 4D face dataset can be applied in the future: 2D/3D digital face generation, face reconstruction capture and other fields.
● Download link: https://opendatalab.org.cn/OpenXD-RenFace/download
No.2 OpenXD-RenBody
● Domain: dynamic 4D human body dataset
● Introduction:
The RendBody dataset covers human body data characteristics of multiple races, age groups, ethnicities, dynasties and special skills. It is currently the largest high-definition multi-view human motion capture dataset. The number of IDs in this dataset reaches 500+, and it captures rich high-definition details of human body shape, movement and clothes in the form of videos with up to 60 viewing angles and 3000x4096 resolution, and provides complete data labeling and processing tools, including personnel information, Manual labeling of clothing types and action categories, as well as tool data annotation such as camera calibration, foreground and background segmentation, and parametric models. Dynamic 4D human body data sets have a wide range of application scenarios in the fields of digital human body reconstruction, animation, and generation.
● Download link: https://opendatalab.org.cn/OpenXD-RenBody/download
No.3 OpenXD-SynBody
● Domain: High-fidelity synthetic human dataset
● Introduction:
The SynBody dataset is a large-scale synthetic human dynamic dataset, which contains rich high-precision character models, action sequences, and various scenes, and uses UnrealEngine to render and output high-fidelity video and annotation information. SynBody builds dynamic scenes of multiple people in different styles of environments. The characters cover more than 700 human body model data of different clothing, body types, genders, and age groups. It uses rich action types such as standing, walking, running, jumping, and dancing to drive the human body model, and provides the corresponding SMPL/SMPLX annotations. This dataset complements the vacancy of large-scale dynamic human synthesis datasets in the academic world, and will support various downstream tasks such as single-view human parametric model estimation, multi-view human parametric model estimation, and person detection and segmentation, and will help promote virtual data Exploration of training methods.
● Download link: https://opendatalab.org.cn/OpenXD-SynBody/download
No.4 OpenXD-OmniObject3D
● Domain: high-quality object scene data
● Introduction:
The Object dataset will build the world's largest high-precision real object 3D dataset, including about 6,000 real object scan models and their surrounding videos, covering 190 common still life categories, filling the vacancy of large-scale real object 3D model datasets in the academic world. This dataset will support the research on the generalization and robustness of various tasks such as neural rendering, surface reconstruction, and point cloud recognition, and will help build a bridge between 2D-3D understanding and algorithm fusion.
● Download link: https://opendatalab.org.cn/OpenXD-OmniObject3D/download
No.5 OpenXD-HuMMan
● Field: multimodal human dataset
● Introduction:
The HuMMan dataset is the world's largest multimodal human dataset, including 1,000 people, 500 movements covering the main muscle groups of the human body, 8 different modalities, more than 400,000 videos, and 60 million frames of data. Data acquisition is based on RGB-D camera and a mobile terminal device, supporting related researches such as action recognition, human body parametric model prediction, and human body surface reconstruction.
● Download link: https://opendatalab.org.cn/OpenXD-HuMMan/download
No.6 OpenXD-GTA-Human
● Domain: Large-Scale Game Generation Datasets
● Introduction:
By cooperating with a large number of computing nodes to run the game Grand Theft Auto V (GTA-V) synchronously, we collected GTA-Human, a large-scale dataset (20,000 videos and 1.4 million frames of SMPL parameter labels). In addition to the scale, GTA-Human uses the rich materials in the game engine to generate a variety of data, some of which are difficult to collect under real conditions: more than 600 characters of different genders, ages, races, body shapes, and clothing; 20,000 different action sequences covering a variety of everyday activities; six main locations, from city streets to the wild, provide completely different backgrounds; camera poses for each sequence are sampled from real-world distributions; human-environment interactions generate different degrees of occlusion; the time in the game affects the lighting conditions; the weather system simulates the climate in reality.
● Download link: https://opendatalab.org.cn/OpenXD-GTA-Human/download
No.7 GeneBody
● Domain: A new dataset for testing the generation of new perspectives on the human body
● Introduction:
GeneBody is a new dataset for testing the generation of new perspectives on the human body. GeneBody provides a total of 370 action sequences of 100 performers with different postures, clothing, and exterior decorations from 48 perspectives, with a total of 2.95 million frames of pictures, covering different human actions and appearances from daily scenes to professional scenes. In addition, the dataset also provides frame-by-frame SMPLx estimation and foreground segmentation.
● Download link: https://opendatalab.org.cn/GeneBody/download
No.8 OpenLane
● Domain: The first real-world and largest 3D lane dataset
● Introduction:
OpenLane is the first real-world and largest 3D lane dataset to date. Our dataset collects valuable content from the public perception dataset Waymo Open Dataset and provides Lane and Closest Path Object (CIPO) annotations for 1000 road segments. In short, OpenLane has 200K frames and over 880K carefully annotated lanes. We made the OpenLane dataset publicly available to help the research community advance 3D perception and autonomous driving technologies.
● Download link: https://opendatalab.org.cn/OpenLane/download
The above is this sharing, for more exciting dry goods of datasets, welcome to visit the official website of OpenDataLab: https://opendatalab.org.cn/ . If there is anything else you want to see, come and tell the little assistant (opendatalab_yunying).
- End -