Link to this sectionTartanAir Depth Dataset#

TartanAir is a large-scale synthetic dataset generated in the AirSim simulator. It was created to push the limits of visual SLAM and spans a wide variety of environments — indoor, outdoor, urban, and nature scenes — along with seasonal, weather, and lighting variation, and challenging conditions.

Because TartanAir is rendered in simulation, it provides dense depth ground truth across this diverse set of scenes, making it a strong source of environmental diversity and long-range geometry for training monocular depth estimation models.

Link to this sectionKey Features#

• Synthetic data generated in the AirSim simulator.
• Diverse indoor and outdoor environments (urban, nature) with seasonal, weather, and lighting variation, plus challenging conditions.
• Dense depth ground truth across all scenes.
• Depth range to ~80 m.
• Contributes 61,470 images (55,660 train / 5,810 val) to the Ultralytics depth training mix.

Link to this sectionDataset Structure#

The TartanAir depth dataset is split into two subsets:

1. Train: 55,660 images with paired dense depth maps for training.
2. Val: 5,810 images with paired dense depth maps for validation during training.

Each RGB image is paired with a .npy float32 depth map storing per-pixel distances in meters, following the Ultralytics depth dataset format.

Link to this sectionObtain the Data#

TartanAir has no autodownload — the data is distributed by CMU's AirLab (see the dataset page for terms) and downloaded with the tartanair_tools scripts:

git clone https://github.com/castacks/tartanair_tools && cd tartanair_tools
python download_training.py --output-dir ./data --rgb --depth --only-left --unzip

Depth is already stored as float32 .npy in meters (depth_left/*_left_depth.npy next to image_left/*_left.png), so conversion is just re-arranging and invalidating the sky (rendered as extreme distances; the released mix clips at 80 m to match the dataset YAML). TartanAir ships no official val split — hold out one or more environments. Reference conversion to the Ultralytics depth dataset format:

import shutil
from pathlib import Path

import numpy as np

VAL_ENVS = {"neighborhood"}  # environments held out for validation
src, dst = Path("data"), Path("datasets/depth-tartanair")
for depth_file in sorted(src.rglob("depth_left/*_left_depth.npy")):
    env, traj = depth_file.parts[-5], depth_file.parts[-3]
    out = "val" if env.lower() in VAL_ENVS else "train"
    (dst / f"images/{out}").mkdir(parents=True, exist_ok=True)
    (dst / f"depth/{out}").mkdir(parents=True, exist_ok=True)
    depth = np.load(depth_file)
    depth[depth > 80.0] = 0.0  # sky/extreme range → 0 = invalid
    frame = depth_file.name.replace("_depth.npy", "")  # e.g. 000000_left
    name = f"{env}_{traj}_{frame}"
    np.save(dst / f"depth/{out}/{name}.npy", depth)
    shutil.copy(depth_file.parents[1] / "image_left" / f"{frame}.png", dst / f"images/{out}/{name}.png")

Link to this sectionRole in YOLO26-Depth#

TartanAir is one of the training sources in the broad multi-dataset mixture (~2.19M images) used to pretrain the Ultralytics YOLO26-Depth models. Within this mix, TartanAir contributes synthetic environmental diversity and long-range outdoor geometry that complement indoor and real-world sources.

There is no standalone held-out TartanAir benchmark in this setup. Instead, the resulting models are evaluated on the standard monocular depth benchmarks: NYU Depth V2, KITTI, Make3D, ETH3D, and iBims-1.

Link to this sectionDataset YAML#

A YAML (Yet Another Markup Language) file is used to define the dataset configuration. It contains information about the dataset's paths, classes, and other relevant information. For TartanAir, the depth-tartanair.yaml file defines the paths and the single depth class.

# Ultralytics 🚀 AGPL-3.0 License - https://ultralytics.com/license

# TartanAir dataset for monocular depth estimation — synthetic indoor + outdoor (AirSim), dense depth up to ~80 m
# Documentation: https://docs.ultralytics.com/datasets/depth/tartanair
# Example usage: yolo depth train data=depth-tartanair.yaml model=yolo26n-depth.pt
# No autodownload — obtain the source data (see docs) and arrange it as below.
# parent
# ├── ultralytics
# └── datasets
#     └── depth-tartanair
#         ├── images/{train,val}  # RGB images
#         └── depth/{train,val}   # paired *.npy, float32 meters (images/ -> depth/)

path: depth-tartanair # dataset root dir (relative to Ultralytics settings 'datasets_dir')
train: images/train # train images (relative to 'path') 55660 images
val: images/val # val images (relative to 'path') 5810 images
max_depth: 80 # (m) maximum valid depth; GT beyond this is excluded from val metrics

nc: 1
names:
  0: depth

channels: 3

Link to this sectionUsage#

To train a YOLO26n-Depth model on the TartanAir dataset with an image size of 640, you can use the following code snippets. For a comprehensive list of available arguments, refer to the model Training page.

from ultralytics import YOLO

# Load a model
model = YOLO("yolo26n-depth.pt")  # load a pretrained model (recommended for training)

# Train the model
results = model.train(data="depth-tartanair.yaml", epochs=100, imgsz=640)

Link to this sectionPretrained Models#

The YOLO26 depth family (yolo26n-depth.pt, yolo26s-depth.pt, yolo26m-depth.pt, yolo26l-depth.pt, yolo26x-depth.pt) auto-downloads from the v8.4.0 release and is trained on the broad multi-dataset mix that TartanAir is part of.

Link to this sectionCitations and Acknowledgments#

If you use the TartanAir dataset in your research or development work, please cite the following paper:

@inproceedings{wang2020tartanair,
      title={TartanAir: A Dataset to Push the Limits of Visual SLAM},
      author={Wenshan Wang and Delong Zhu and Xiangwei Wang and Yaoyu Hu and Yuheng Qiu and Chen Wang and Yafei Hu and Ashish Kapoor and Sebastian Scherer},
      booktitle={IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)},
      year={2020}
}

We would like to acknowledge the creators of TartanAir for making this diverse synthetic dataset available to the computer vision community.