nvidia/PhysicalAI-SmartSpaces

Physical AI Smart Spaces Dataset

Overview

Comprehensive, annotated dataset for multi-camera tracking and 2D/3D object detection. This dataset is synthetically generated with Omniverse.

This dataset consists of over 250 hours of video from across nearly 1,500 cameras from indoor scenes in warehouses, hospitals, retail, and more. The dataset is time synchronized for tracking humans across multiple cameras using feature representation and no personal data.

Dataset Description

Dataset Owner(s)

NVIDIA

Dataset Creation Date

We started to create this dataset in December, 2023. First version was completed and released as part of 8th AI City Challenge in conjunction with CVPR 2024.

Dataset Characterization

• Data Collection Method: Synthetic
• Labeling Method: Automatic with IsaacSim

Video Format

• Video Standard: MP4 (H.264)
• Video Resolution: 1080p
• Video Frame rate: 30 FPS

Ground Truth Format (MOTChallenge) for MTMC_Tracking_2024

Annotations are provided in the following text format per line:

<camera_id> <obj_id> <frame_id> <xmin> <ymin> <width> <height> <xworld> <yworld>

• <camera_id>: Numeric identifier for the camera.
• <obj_id>: Consistent numeric identifier for each object across cameras.
• <frame_id>: Frame index starting from 0.
• <xmin> <ymin> <width> <height>: Axis-aligned bounding box coordinates in pixels (top-left origin).
• <xworld> <yworld>: Global coordinates (projected bottom points of objects) based on provided camera matrices.

The video file and calibration (camera matrix and homography) are provided for each camera view.

Directory Structure for MTMC_Tracking_2025

• videos/: Video files.
• ground_truth.json: Detailed ground truth annotations (see below).
• calibration.json: Camera calibration and metadata.
• map.png: Visualization map in top-down view.

Ground Truth Format (JSON) for MTMC_Tracking_2025

Annotations per frame:

{
  "<frame_id>": [
    {
      "object_type": "<class_name>",
      "object_id": <int>,
      "3d_location": [x, y, z],
      "3d_bounding_box_scale": [w, l, h],
      "3d_bounding_box_rotation": [pitch, roll, yaw],
      "2d_bounding_box_visible": {
        "<camera_id>": [xmin, ymin, xmax, ymax]
      }
    }
  ]
}

Calibration Format (JSON) for MTMC_Tracking_2025

Contains detailed calibration metadata per sensor:

{
  "calibrationType": "cartesian",
  "sensors": [
    {
      "type": "camera",
      "id": "<sensor_id>",
      "coordinates": {"x": float, "y": float},
      "scaleFactor": float,
      "translationToGlobalCoordinates": {"x": float, "y": float},
      "attributes": [
        {"name": "fps", "value": float},
        {"name": "direction", "value": float},
        {"name": "direction3d", "value": "float,float,float"},
        {"name": "frameWidth", "value": int},
        {"name": "frameHeight", "value": int}
      ],
      "intrinsicMatrix": [[f_x, 0, c_x], [0, f_y, c_y], [0, 0, 1]],
      "extrinsicMatrix": [[3×4 matrix]],
      "cameraMatrix": [[3×4 matrix]],
      "homography": [[3×3 matrix]]
    }
  ]
}

Evaluation

• 2024 Edition: Evaluation based on HOTA scores at the 2024 AI City Challenge Server. The submission is currently disabled, as the ground truths of test set are provided with this release.
• 2025 Edition: Evaluation system and test set forthcoming in the 2025 AI City Challenge.

Dataset Quantification

Dataset	Annotation Type	Hours	Cameras	Object Classes & Counts	No. 3D Boxes	No. 2D Boxes	Total Size
MTMC_Tracking_2024	2D bounding boxes, multi-camera tracking IDs	212	953	Person: 2,481	N/A	135M	198 GB
MTMC_Tracking_2025 (Train & Validation only)	2D & 3D bounding boxes, multi-camera tracking IDs	42	504	Person: 292 Forklift: 13 NovaCarter: 28 Transporter: 23 FourierGR1T2: 6 AgilityDigit: 1 Overall: 363	8.9M	73M	74 GB

References

Please cite the following papers when using this dataset:

@InProceedings{Wang24AICity24,
author = {Shuo Wang and David C. Anastasiu and Zheng Tang and Ming-Ching Chang and Yue Yao and Liang Zheng and Mohammed Shaiqur Rahman and Meenakshi S. Arya and Anuj Sharma and Pranamesh Chakraborty and Sanjita Prajapati and Quan Kong and Norimasa Kobori and Munkhjargal Gochoo and Munkh-Erdene Otgonbold and Ganzorig Batnasan and Fady Alnajjar and Ping-Yang Chen and Jun-Wei Hsieh and Xunlei Wu and Sameer Satish Pusegaonkar and Yizhou Wang and Sujit Biswas and Rama Chellappa},
title = {The 8th {AI City Challenge},
booktitle = {The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops},
note = {arXiv:2404.09432},
month = {June},
year = {2024},
}

@misc{Wang24BEVSUSHI,
author = {Yizhou Wang and Tim Meinhardt and Orcun Cetintas and Cheng-Yen Yang and Sameer Satish Pusegaonkar and Benjamin Missaoui and Sujit Biswas and Zheng Tang and Laura Leal-Taix{\'e}},
title = {{BEV-SUSHI}: {M}ulti-target multi-camera {3D} detection and tracking in bird's-eye view},
note = {arXiv:2412.00692},
year = {2024}
}

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 model team to ensure this model meets requirements for the relevant industry and use case and addresses unforeseen product misuse.

Please report security vulnerabilities or NVIDIA AI Concerns here.