PDF Paper ArXiv arXiv Youtube Video Github

AirSim360

A Panoramic Simulation Platform within Drone View


Xian Ge, Yuling Pan, Yuhang Zhang, Xiang Li, Weijun Zhang, Dizhe Zhang,
Zhaoliang Wan, Xin Lin, Xiangkai Zhang, Juntao Liang,
Jason Li, Wenjie Jiang, Bo Du, Ming-Hsuan Yang, Lu Qi


Insta360 Logo

PDF Paper ArXiv arXiv Code

Technical Summary Video

AirSim360, a simulation platform for omnidirectional data from aerial viewpoints, enabling wide-ranging scene sampling with drones. Specifically, AirSim360 focuses on three key aspects: a render-aligned data and labeling paradigm for pixel-level geometric, semantic, and instance-level understanding; an interactive pedestrian-aware system for modeling human behavior; and an automated trajectory generation paradigm to support navigation tasks.

AirSim360 Architecture

AirSim360 is a large-scale omnidirectional simulation platform for drone scenarios. we emphasize consistency and compatibility among three key components in low-altitude ground environments, including static surroundings, the UAV, and human actors, which aligns closely with the main characteristics of the human-centric real world. Projection figure

Datasets Collection

To serve diverse research objectives, Omni360-X is organized into three subsets, including Scene, Pedestrian, and Trajectory, where each emphasizes a specific aspect of 360-degree understanding. Gaps figure Gaps figure

Experiments

We present and benchmark the sampled dataset in multiple tasks, indicating the benefits of our proposed simulator for real-world scenarios. Our simulated data consistently improves performance and robustness when evaluated on public real-world validation sets.

1. Monocular Pedestrian Distance Estimation (MPDE)

Training with the Omni360-Human dataset significantly improves the cross-domain generalization of Monocular Pedestrian Distance Estimation (MPDE) models. Specifically, the average angular error decreases from 21.21° to 17.02°, and the mean distance error decreases from 0.484 m to 0.458 m on public test sets.

Training Set Test Set Dist. Err ↓ (m) Ang. Err ↓ (°)
nuScenes KITTI 0.822 31.50
nuScenes + Omni360 KITTI 0.809 31.20
nuScenes FreeMan 0.260 17.00
nuScenes + Omni360 FreeMan 0.228 11.60

2. Panoramic Depth Estimation (Out-of-Domain)

We conduct out-of-domain experiments to evaluate the generalization ability of the Omni360 dataset. Models trained on Omni360 show improved generalization to unseen environments (SphereCraft).

Experiment Type Training Data Test Data AbsRel ↓ RMSE ↓
Out-of-Domain Deep360 SphereCraft 8.2570 0.0566
Out-of-Domain Omni360 SphereCraft 5.4372 0.0435

3. Panoramic Segmentation

The introduction of our large-scale Omni360-Scene segmentation data provides substantial performance gains on the WildPASS validation set for pixel-level scene parsing.

Task Training Data Performance Metric Result
Semantic Segmentation WildPASS Baseline mIoU ↑ 58.0
WildPASS + Omni360 67.4
Entity Segmentation WildPASS Baseline mAP ↑ 24.6
WildPASS + Omni360 38.9

Citation

@article{ge2025airsim360,
  title={AirSim360: A Panoramic Simulation Platform within Drone View },
  author={Ge, Xian and Pan, Yuling and Zhang, Yuhang and Li, Xiang and Zhang, Weijun and Zhang, Dizhe and Wan, Zhaoliang and Lin, Xin and Zhang, Xiangkai and Liang, Juntao and Li, Jason and Jiang, Wenjie and Du, Bo and Yang, Ming-Hsuan and Qi, Lu},
  journal={arXiv},
  year={2025}
}