--- license: cc-by-4.0 configs: - config_name: actions data_files: - split: full path: - annotation_files/actions.csv format: csv - config_name: bounding_boxes data_files: - split: full path: - annotation_files/bounding_boxes.csv format: csv - config_name: vqa_and_captioning data_files: - split: full path: - vqa_and_captioning/vqa_and_captioning.json format: json language: - en tags: - human action recognition - video understanding - microgravity task_categories: - video-classification - visual-question-answering - video-text-to-text --- # MicroG-4M Dataset This repository stores the entire content of the MicroG-4M dataset itself. **For more information and details, including training, evaluation, statistics, and related code, please:** - **Refer to our [paper](https://arxiv.org/abs/2506.02845)** - **Visit our [GitHub](https://github.com/LEI-QI-233/MicroG-4M)** - **And check our [fine-tuned models](https://huggingface.co/LEI-QI-233/MicroG-4M-models)** # Specification of MicroG-4M ## "annotation_files" Folder The folder contains all annotation files of the dataset, all stored in CSV format. ### [actions.csv](./annotation_files/actions.csv) contains all the labeled persons in the video clips and their corresponding action IDs. The headers are: - `video_id` : is the [name](#video-name) of video clip. - `movie_or_real`: `m` for movie and `r` for real footage - `person_id`: corresponds to the person labeled by the respective bounding box. You can visually identify the corresponding person directly in the videos within the [videos_annotated](./videos_annotated.zip) folder. - `action`: the ID of action. For the mapping relationship between action IDs and labels see [label_map](#label_map-folder) ### [bounding_boxes.csv](./annotation_files/bounding_boxes.csv) includes all detected persons and their corresponding bounding box coordinates. The coordinates are expressed in pixels rather than normalized values. The headers are: - `video_id`: is the [name](#video-name) of video clip. - `frame_id`: Not a continuous number. Only frames where the bounding box is drawn. If no person is detected, ignore this frame. - `person_id` - `xmin`: x coordinate of the upper left corner - `ymin`: y coordinate of the upper left corner - `xmax`: x coordinate of the lower right corner - `ymax`: y coordinate of the lower right corner ### NOTE Each row in the [actions.csv](./annotation_files/actions.csv) and [bounding_boxes.csv](./annotation_files/bounding_boxes.csv) files is associated through the video ID and person ID, forming a complete information structure that includes: - video ID - frame ID - person ID - movie or real - action ID - bounding box coordinates ## "label_map" Folder The folder contains the pbtxt file of the label map as well as a PDF file. ### [label_map.pbtxt](./label_map/label_map.pbtxt) is the pbtxt format file of the label map, containing three components: id, name, and type. Its format is consistent with the label map format of the [AVA](https://research.google.com/ava/) dataset. If this dataset is converted to the AVA format for training or evaluation, this label map file can be directly used as input. ### [label_map.pdf](./label_map/label_map.pdf) is the PDF format of the label map, providing a convenient reference for users to read and consult. ## "vqa_and_captioning" Folder The folder contains the json files for Visual Queastion Qnswering and captioning. ### [vqa_and_captioning.json](./vqa_and_captioning/vqa_and_captioning.json) is an array containing information for multiple videos, with each record consisting of the following parts: - `video_id`: The ID of the video - `caption`: A brief description of the video content, used for generating subtitles - `qa_pairs`: A list of question-answer pairs, where each object includes two fields: - `Q`: A question related to the video scene or context - `A`: The corresponding answer to the question ## "videos" Folder The [`video_id_list.pdf`](./videos/video_id_list.pdf) file in this folder provides the identifiers for all videos used in the dataset. **Users should download the raw files independently** and process them into the required format as described in our paper. You can process the downloaded videos into our dataset format using your own tools or our provided automated pipeline, please refer to [`CODE.md`](https://github.com/lei-qi-233/MicroG-4M/blob/main/code/CODE.md) in our GitHub repository. ## Video Specification ### Format of processed videos All videos are processed to video clips which are 3 seconds long, with a frame rate of 30 fps and encoded in H.264 format. You can find detailed information in our Github and our paper. ### Video Folder Structure The video folder should has the following structure: ``` videos |_ movie | |_ [movie name 0] | | |_ [movie name 0]_000.mp4 | | |_ [movie name 0]_003.mp4 | | |_ ... | |_ [movie name 1] | | |_ ... |_ real | |_ [real video id 0] | | |_[real video id 0]_002.mp4 | | |_[real video id 0]_003.mp4 | | |_ ... | |_ [real video id 1] | | |_ ... ``` #### Video Name The video name consists of a sequence of characters and underscores followed by a sequence of numbers, and corresponds to the `video_id` column in the CSV and JSON annotation files. The character string is either the YouTube video ID or the film name. The sequence number at the end of the video name indicates its specific position within the original video. For example, 000 represents the 0 to 3 seconds of the original video, and 003 represents the 9 to 12 seconds. Since not all clips meet the criteria, the sequence numbers of video clips in the dataset are not consecutive. For example: - `AErpXJq67LM_000` represents the first three-second segment of the YouTube video with ID `AErpXJq67LM` - `Apollo13_101` denotes the segment of the movie `Apollo13` spanning the 303rd to the 306rd second (i.e., from number × 3 to (number × 3) + 3 seconds). #### `movie` and `real` folders - In the `movie` folder, all video clips are from movies. `movie name` is the name of the movie. - In the `real` folder, all videos are real footage and downloaded from YouTube. `real video id` is the video id on YouTube. ### For further information, please refer to our paper. # Citing MicroG-4M If you find this dataset useful in your research, please use the following BibTeX entry for citation: ```bibtex @inproceedings{wen2026go, title={{Go Beyond Earth: Understanding Human Actions and Scenes in Microgravity Environments}}, author={Di Wen and Lei Qi and Kunyu Peng and Kailun Yang and Fei Teng and Ao Luo and Jia Fu and Yufan Chen and Ruiping Liu and Yitian Shi and M. Saquib Sarfraz and Rainer Stiefelhagen}, booktitle={International Conference on Learning Representations (ICLR)}, year={2026}, url={https://openreview.net/forum?id=gygGCVXeh3} } ```