MCG-NJU/VideoChatOnline-IT

Overview

This dataset provides a comprehensive collection for Online Spatial-Temporal Understanding tasks, covering multiple domains including Dense Video Captioning, Video Grounding, Step Localization, Spatial-Temporal Action Localization, and Object Tracking.

Data Formation

Our pipeline begins with 96K high-quality samples curated from 5 tasks across 12 datasets. The conversion process enhances online spatiotemporal understanding through template transformation. We strategically insert queries along the timeline in an organized interleaved format for each video sample to facilitate temporal context differentiation.

Category	Dataset	Count	Query	Response
Temporal Grounding	DiDeMo	33,002	Identify whether a specific event is still ongoing at present or has it concluded. Provide the start time of the event and its duration up to the query timestamp.	<start time> - <event duration>: duration up to query timestamp.
	QuerYD	14,620
	HiREST	459
	Charades-STA	12,408
Object Tracking	LaSOT	1,120	Track the object currently based on a brief description or box.	(1) Past trajectory up to the present with brief descriptions; (2) Track the object sequentially in future frames as they become available.
	GOT10k	8,250
Step Localization and Captioning	COIN	9,029	List steps completed up to the current point, excluding previously reported ones.	<start time> - <end time>, <step description>...
	HiREST	459
Dense Video Captioning	ActivityNet Captions	10,009	Identify and list events up to the current point, excluding previously reported ones.	<start time> - <end time>, <event description>...
	VITT	5,141
	YouCook2	1,192
Spatial Temporal Action Localization	AVA	160	Identify current and past actions of a person at a specific box at present.	List actions for the person over time, with corresponding positions.
Total number of datasets:		96k

---

Additional Information:

• Interleave Format: Temporally Random Insert (T3, T2, T1)
• Video Timeline: Processed for Online Video LLM

Data Formats

• Format 1: Conversational QA (LLaVA-style)

{
    "video": "116/NLy71UrHElw.mp4",
    "conversations": [
        {
            "from": "human",
            "timestamps": 1026.0,  # Video timestamp in seconds
            "value": "<video>\nBased on current observation, list events..."
        },
        {
            "from": "gpt",
            "value": "21.0s - 22.0s (duration: 1.0s), begin to run up..."
        }
    ]
}

Format 2: Template-based Tracking

{
    "video": "GOT-10k_Train_006457",
    "fps": 1,  # Frame rate
    "all_image_files": ["00000001.jpg", ...],  # Keyframe paths
    "image_bboxes": [  # Temporal object tracking data
        {
            "timestamp": 0.0,
            "bbox": [0.412, 0.517, 0.452, 0.753]  # [x1,y1,x2,y2]
        },
        ...
    ],
    "query_template": {  # Randomized temporal insertion
        "from": "human",
        "value": "Track the location of \"person\" at <bbox> over time..."
    }
}

Source Data

Task	Dataset	Source
Dense Video Captioning	ActivityNet Captions	Source
	ViTT	Source
	YouCook2	Source
Temporal Video Grounding	DiDeMo	Source
	QuerYD	Source
	HiREST_grounding	Source
	Charades-STA	Source
Step Localization	COIN	Source
	HiREST_step	Source
Spatial Temporal Action Localization	AVA	Source
Object Tracking	GOT 10K	Source
	LaSOT	Source

Citation

If you find this project useful in your research, please consider cite:

@article{huang2024online,
  title={Online Video Understanding: A Comprehensive Benchmark and Memory-Augmented Method},
  author={Huang, Zhenpeng and Li, Xinhao and Li, Jiaqi and Wang, Jing and Zeng, Xiangyu and Liang, Cheng and Wu, Tao and Chen, Xi and Li, Liang and Wang, Limin},
  journal={arXiv preprint arXiv:2501.00584},
  year={2024}
}