nvidia/Kimodo-Motion-Gen-Benchmark
收藏资源简介:
--- license: cc-by-4.0 language: - en tags: - Kimodo - nvidia - human motion generation pretty_name: Kimodo Human Motion Generation Benchmark viewer: false --- # Kimodo Human Motion Generation Benchmark [Kimodo Codebase](https://github.com/nv-tlabs/kimodo), [Benchmark Documentation](https://research.nvidia.com/labs/sil/projects/kimodo/docs/benchmark/introduction.html) ## Dataset Description: This dataset provides the necessary metadata to construct the suite of test cases that make up the [Kimodo](https://github.com/nv-tlabs/kimodo) human motion generation benchmark. This includes test cases that evaluate text-following for the text-to-motion task, along with constraint-following for constraint-conditioned motion generation. The benchmark is constructed from the SOMA uniform version of the [BONES-SEED dataset](https://huggingface.co/datasets/bones-studio/seed) using metadata included in this repo, which includes text prompts derived from our [SEED timeline annotations](https://huggingface.co/datasets/nvidia/SEED-Timeline-Annotations), start/end frames for test case motions, and pose constraint configurations. To get started with the benchmark, see the [Kimodo documentation](https://research.nvidia.com/labs/sil/projects/kimodo/docs/benchmark/introduction.html). This repo also includes the train and test splits for the BONES-SEED data that should be used to train models that will be evaluated with the Kimodo benchmark. This dataset is ready for commercial use. ## Dataset Owner: NVIDIA Corporation ## Dataset Creation Date: April 2026 ## License/Terms of Use: This dataset is governed by the [Creative Commons Attribution 4.0 International License](https://creativecommons.org/licenses/by/4.0/) (CC BY 4.0). ## Intended Usage: This dataset is intended for researchers and developers training motion generation models on BONES-SEED to evaluate their models in a comprehensive and standardized way. ## Download Instructions: The easiest way to download the dataset is using Git: ``` git clone git@hf.co:datasets/nvidia/Kimodo-Motion-Gen-Benchmark ``` ## Dataset Structure: For full details of the benchmark, please see the [Kimodo documentation](https://research.nvidia.com/labs/sil/projects/kimodo/docs/benchmark/introduction.html). Train and test splits are defined in `splits`: - `train_split_paths.txt` - filenames of training data - `test_content_split_paths.txt` - filenames for test split containing new semantic "content". This split contains motions with `content_name` (from the BONES-SEED metadata) that are not seen in the training split. This tests model generalization to new semantic motion types, e.g. for text-to-motion generalization. - `test_repetition_split_paths.txt` - filenames for test split containing new motions from content that was seen in training. This split contains motions where the `content_name` is contained in the training split, but the exact motion itself was not seen. This tests a model's ability to generalize to novel performances of a familiar motion type, e.g., for constraint-following generalization. The benchmark test cases are partitioned into the `content` and `repetition` test splits. Note that the test cases in the benchmark cover a diverse subset of these splits. Within each of the split directories, the structure categorizes test cases into tasks ranging from pure text-to-motion to constraint-conditioned generation with a variety of different constraint types and scenarios: ```text testsuite ├── content │ ├── constraints_notext │ │ ├── end-effectors │ │ ├── fullbody │ │ ├── mixture │ │ └── root │ ├── constraints_withtext │ │ ├── end-effectors │ │ ├── fullbody │ │ ├── mixture │ │ └── root │ └── text2motion │ ├── overview │ ├── timeline_multi │ └── timeline_single └── repetition ├── constraints_notext │ ├── end-effectors │ ├── fullbody │ ├── mixture │ └── root ├── constraints_withtext │ ├── end-effectors │ ├── fullbody │ ├── mixture │ └── root └── text2motion ├── overview ├── timeline_multi └── timeline_single ``` At the lowest level of this structure, each leaf folder contains indexed test cases (`0000`, `0001`, `0002`, ...). For example: ```text end-effectors/feet_posrot/ ├── 0000/ ├── 0001/ ├── 0002/ ... └── 0255/ ``` Each index folder is one standalone test case with its own `meta.json` containing text prompt and duration, `seed_motion.json` with metadata relating the test case to BONES-SEED, and optionally `seed_constraints.json` defining constraints for the test case. These are used to process the BONES-SEED dataset to build the full benchmark. All text prompts in the benchmark are derived from our [SEED timeline annotations](https://huggingface.co/datasets/nvidia/SEED-Timeline-Annotations). ## Dataset Quantification: - Dataset file size: 116 MB - Total number of test cases: 22,474 ## References: - [Kimodo Motion Generation Model](https://research.nvidia.com/labs/sil/projects/kimodo/) - [BONES-SEED dataset](https://huggingface.co/datasets/bones-studio/seed) ## Ethical Considerations: NVIDIA believes Trustworthy AI is a shared responsibility and we have established policies and practices to enable development for a wide array of AI applications. When downloaded or used in accordance with our terms of service, developers should work with their internal developer teams to ensure this dataset meets requirements for the relevant industry and use case and addresses unforeseen product misuse. Please report quality, risk, security vulnerabilities or NVIDIA AI Concerns [here](https://app.intigriti.com/programs/nvidia/nvidiavdp/detail).
--- 许可证:CC BY 4.0 语言: - 英语 标签: - Kimodo - 英伟达(NVIDIA) - 人体运动生成 美观名称:Kimodo人体运动生成基准(Kimodo Human Motion Generation Benchmark) 查看器:禁用 --- # Kimodo人体运动生成基准(Kimodo Human Motion Generation Benchmark) [Kimodo代码库](https://github.com/nv-tlabs/kimodo)、[基准测试文档](https://research.nvidia.com/labs/sil/projects/kimodo/docs/benchmark/introduction.html) ## 数据集描述: 本数据集提供构建Kimodo人体运动生成基准(Kimodo Human Motion Generation Benchmark)所需的全套测试用例元数据。其中包含用于评估文本到运动(text-to-motion)任务文本遵循能力的测试用例,以及用于约束条件运动生成的约束遵循能力测试用例。 该基准基于BONES-SEED数据集(BONES-SEED dataset)的SOMA统一版本构建,使用本仓库中包含的元数据完成构建。这些元数据包括源自我们的SEED时间线标注(SEED Timeline Annotations)的文本提示词、测试用例运动的起始/结束帧,以及姿态约束配置。 若要开始使用该基准,请参阅Kimodo官方文档。 本仓库还包含BONES-SEED数据集的训练集与测试集划分信息,供需要在Kimodo基准上进行模型评估的训练任务使用。 本数据集可直接用于商业用途。 ## 数据集所有者: 英伟达公司(NVIDIA Corporation) ## 数据集创建日期: 2026年4月 ## 许可证与使用条款: 本数据集受知识共享署名4.0国际许可协议(Creative Commons Attribution 4.0 International License,简称CC BY 4.0)管辖。 ## 预期用途: 本数据集面向基于BONES-SEED数据集训练运动生成模型的研究人员与开发者,用于以全面且标准化的方式评估其模型性能。 ## 下载指南: 下载本数据集最简方式为使用Git克隆: git clone git@hf.co:datasets/nvidia/Kimodo-Motion-Gen-Benchmark ## 数据集结构: 若需了解基准的完整细节,请参阅Kimodo官方文档。 训练集与测试集划分信息存储于`splits`目录下: - `train_split_paths.txt`:训练数据文件名列表 - `test_content_split_paths.txt`:包含全新语义“内容”的测试集文件名列表。该测试集包含训练集中未出现的`content_name`(源自BONES-SEED元数据)对应的运动数据,用于评估模型对全新语义运动类型的泛化能力,例如文本到运动任务的泛化性能。 - `test_repetition_split_paths.txt`:基于训练集中已有内容生成全新运动的测试集文件名列表。该测试集包含`content_name`已出现在训练集中,但具体运动样本未在训练集中出现的数据,用于评估模型对熟悉运动类型的新颖表现的泛化能力,例如约束遵循任务的泛化性能。 基准测试用例被划分为`content`与`repetition`两类测试集。需注意,基准中的测试用例涵盖了这两类划分下的多样化子集。在每个划分目录下,测试用例按任务类型进行分类,覆盖从纯文本到运动生成到结合多种约束类型与场景的约束条件运动生成任务: text testsuite ├── content │ ├── constraints_notext │ │ ├── end-effectors │ │ ├── fullbody │ │ ├── mixture │ │ └── root │ ├── constraints_withtext │ │ ├── end-effectors │ │ ├── fullbody │ │ ├── mixture │ │ └── root │ └── text2motion │ ├── overview │ ├── timeline_multi │ └── timeline_single └── repetition ├── constraints_notext │ ├── end-effectors │ ├── fullbody │ ├── mixture │ └── root ├── constraints_withtext │ ├── end-effectors │ ├── fullbody │ ├── mixture │ └── root └── text2motion ├── overview ├── timeline_multi └── timeline_single 在该结构的最底层,每个叶节点文件夹均包含带索引的测试用例(`0000`、`0001`、`0002`……)。示例如下: text end-effectors/feet_posrot/ ├── 0000/ ├── 0001/ ├── 0002/ ... └── 0255/ 每个索引文件夹对应一个独立测试用例,包含专属的`meta.json`(存储文本提示词与运动时长)、`seed_motion.json`(记录测试用例与BONES-SEED数据集关联的元数据),以及可选的`seed_constraints.json`(定义测试用例的约束配置)。这些文件用于处理BONES-SEED数据集以构建完整的基准测试套件。 本基准中的所有文本提示词均源自我们的SEED时间线标注数据集。 ## 数据集量化统计: - 数据集文件大小:116 MB - 测试用例总数:22474个 ## 参考文献: - [Kimodo运动生成模型](https://research.nvidia.com/labs/sil/projects/kimodo/) - [BONES-SEED数据集](https://huggingface.co/datasets/bones-studio/seed) ## 伦理考量: 英伟达坚信,可信人工智能(Trustworthy AI)是一项共同责任,我们已建立相关政策与实践规范,以支持各类人工智能应用的开发。开发者在按照本服务条款下载或使用本数据集时,应与其内部开发团队协作,确保本数据集符合相关行业与使用场景的要求,并规避可能出现的产品误用问题。 若需报告质量问题、风险、安全漏洞或英伟达人工智能相关问题,请访问[此处](https://app.intigriti.com/programs/nvidia/nvidiavdp/detail).




