ImageNet (Pseudo-Labeled) Depth Dataset

The ImageNet (pseudo-labeled depth) dataset reuses the 1,281,167 ImageNet-1K training images, which have no ground-truth depth, and pairs each image with a pseudo depth map produced by a Depth Anything 3 teacher, combining its monocular and metric checkpoints. It is used purely for knowledge distillation and for the scene and object diversity that ImageNet provides. As the single largest source in the roughly 2.19M-image YOLO26-Depth pretraining mix (about 58%), adding it was decisive for indoor cross-domain generalization.

Key Features

1,281,167 ImageNet-1K training images covering a very broad range of objects, scenes, and contexts.
No ground-truth depth: every depth target is a pseudo label generated by a Depth Anything 3 teacher pairing (DA3MONO-LARGE for scene structure, DA3METRIC-LARGE for metric scale).
Used purely for knowledge distillation and scene/object diversity, not as a benchmark.
The single largest source in the ~2.19M-image pretraining mix (~58%), decisive for indoor cross-domain generalization.
No validation split — validation is performed only on real ground-truth datasets.

Dataset Structure

The pseudo-labeled ImageNet source consists of the ImageNet-1K training images with one machine-generated depth map per image:

Training images: 1,281,167 ImageNet-1K training images, the largest single component (~58%) of the ~2.19M-image YOLO26-Depth pretraining mix.
Validation: none. This source has no validation split; YOLO26-Depth validation uses only real ground-truth datasets such as NYU Depth V2 and KITTI Eigen.

How the pseudo-labels are generated

Because ImageNet ships without depth annotations, depth targets are produced offline rather than measured. Two Depth Anything 3 checkpoints are combined, one for scene structure and one for absolute scale:

Each ImageNet training image is run through depth-anything/DA3MONO-LARGE, which predicts a highly detailed depth map that is only defined up to an unknown per-image scale and shift, and through depth-anything/DA3METRIC-LARGE, whose output is coarser but in real units.
A per-image robust affine fit maps the monocular prediction onto the metric one, label = a * mono + b. Every per-pixel detail therefore comes from the monocular checkpoint, while the metric checkpoint only sets the two scalars that place the map on the meter axis. No camera intrinsics are involved, so images without calibration can be labeled.
The result is saved as a .npy array in meters, following the Ultralytics depth dataset format, and paired with its source image by file stem. These maps are stored as float16 to halve the on-disk footprint across 1.28M files; the dataset loader casts them to float32 on read, and float32 remains the documented format for custom depth datasets.
The pseudo-labeled pairs are then added as one component of a larger training mix, where a student YOLO26-Depth model learns to match the teacher while also training on real ground-truth sources.

Because the scale comes from a prediction rather than a measurement, each map can carry a global scale error. YOLO26-Depth trains with a scale-invariant log (SILog) loss plus gradient matching and validates with median alignment, so a per-image scale offset in the pseudo labels is largely absorbed.

This pseudo-labeled source is a component of the internal mixed-training configuration used to pretrain the released YOLO26-Depth weights and is not distributed as a standalone download.

Role in YOLO26-Depth

This source contributes the bulk of the YOLO26-Depth pretraining data and the widest diversity of scenes and objects. By distilling a strong Depth Anything 3 teacher across more than a million images, it transfers broad scene priors into the student. In practice, adding it was decisive for indoor cross-domain generalization, helping the model perform well on indoor scenes beyond the real ground-truth training distributions.

Usage

The pseudo-labeled ImageNet data is not trained on in isolation; it is consumed as one component of a combined depth mix, defined by an internal/experiment YAML rather than a public dataset download. Conceptually, training on such a mix looks like the following:

!!! example "Train Example"

=== "Python"

    ```python
    from ultralytics import YOLO

    # Load a pretrained depth model
    model = YOLO("yolo26n-depth.pt")

    # Train on a combined depth mix (your own multi-source dataset YAML)
    results = model.train(data="depth-mix.yaml", epochs=100, imgsz=640)
    ```

=== "CLI"

    ```bash
    # Train on a combined depth mix (your own multi-source dataset YAML)
    yolo depth train data=depth-mix.yaml model=yolo26n-depth.pt epochs=100 imgsz=640
    ```

Pretrained Models

Pretrained YOLO26-Depth models auto-download from the Ultralytics v8.4.0 assets release when first referenced by name:

Citations and Acknowledgments

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

!!! quote ""

=== "BibTeX"

    ```bibtex
    @inproceedings{deng2009imagenet,
          title={ImageNet: A Large-Scale Hierarchical Image Database},
          author={Deng, Jia and Dong, Wei and Socher, Richard and Li, Li-Jia and Li, Kai and Fei-Fei, Li},
          booktitle={2009 IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
          year={2009},
          organization={IEEE}
    }

    @article{depthanything3,
          title={Depth Anything 3: Recovering the visual space from any views},
          author={Lin, Haotong and Chen, Sili and Liew, Jun Hao and Chen, Donny Y. and Li, Zhenyu and Shi, Guang and Feng, Jiashi and Kang, Bingyi},
          journal={arXiv preprint arXiv:2511.10647},
          year={2025}
    }
    ```

We would like to acknowledge the ImageNet team for creating and maintaining the dataset, and the authors of Depth Anything 3 for the teacher models used to generate the pseudo depth labels.