supermemory-vqa-imu-benchmark
收藏资源简介:
SuperMemory-VQA 是一个面向流式第一人称视频与 IMU 数据的视觉问答基准测试清单。该数据集作为 `kfkas/supermemory-vqa-imu-benchmark` 的查看器就绪清单,包含基准标注和本地资产提示,但不重新分发原始媒体文件。基础数据集为 SuperMemory-VQA Aria 流,共有 4,853 行,其中 4,818 行同时包含本地 RGB 视频和 IMU 数据。所有答案密钥均为公开状态。所有行使用统一的列结构,包括:`benchmark`、`base_dataset`、`question_id`、`category`、`question`、`choices`、`answer_index`、`answer_text`、`source_ids`、`query_time`、`evidence_time`、`clip_start_s`、`clip_end_s`、`stream_count`、`video_asset_hint`、`imu_asset_hint`、`local_video_available`、`local_imu_available`、`usable_for_imu_benchmark`、`coverage_status`、`answer_key_status`、`metadata_json` 和 `streams`。`streams` 字段采用一致的嵌套结构,包含源 ID、模态、传感器角色、时间边界、资产提示、本地可用性及连接类型。数据集特定字段编码在 `metadata_json` 中,以保持查看器模式稳定。用户可通过 Hugging Face datasets 库加载数据集,并使用 `filter` 选择适用于 IMU 基准测试的样本。数据集不包含完整的原始 MP4、VRS、ZIP、CSV 或 NPZ 文件,仅提供一段短示例视频和相对路径提示,完整源媒体需从上游获取。
SuperMemory-VQA is a visual question answering benchmark for streaming first-person video and IMU data. It serves as a viewer-ready checklist for the `kfkas/supermemory-vqa-imu-benchmark` dataset, containing benchmark annotations and local asset hints, but does not redistribute original media files. The base dataset is SuperMemory-VQA Aria Stream, with 4,853 rows, of which 4,818 include both local RGB video and IMU data. All answer keys are public. All rows share a unified column structure including: `benchmark`, `base_dataset`, `question_id`, `category`, `question`, `choices`, `answer_index`, `answer_text`, `source_ids`, `query_time`, `evidence_time`, `clip_start_s`, `clip_end_s`, `stream_count`, `video_asset_hint`, `imu_asset_hint`, `local_video_available`, `local_imu_available`, `usable_for_imu_benchmark`, `coverage_status`, `answer_key_status`, `metadata_json`, and `streams`. The `streams` field has a consistent nested structure containing source ID, modality, sensor role, temporal boundaries, asset hints, local availability, and connection type. Dataset-specific fields are encoded in `metadata_json` to maintain viewer mode stability. Users can load the dataset via the Hugging Face datasets library and use `filter` to select samples suitable for IMU benchmarks. The dataset does not contain complete original MP4, VRS, ZIP, CSV, or NPZ files; it only provides a short example video and relative path hints. Full source media must be obtained from upstream sources.





