Awesome

Depth Anything ONNX

Open Neural Network Exchange (ONNX) compatible implementation of Depth Anything: Unleashing the Power of Large-Scale Unlabeled Data and Depth Anything V2. Supports PyTorch 2 Export via TorchDynamo.

<img src="assets/demo.gif" autoplay width=90%> Monocular Depth Estimation with Depth Anything V2 <img src="assets/sample.png" width=90%>

Changelog:

22 June 2024: Support Depth Anything V2 and TorchDynamo Export.
22 January 2024: Release.

🔥 ONNX Export & Inference

We provide a simple command-line tool dynamo.py based on Typer to export Depth Anything V2 to ONNX and PyTorch2 programs. Please install the requirements first.

$ python dynamo.py --help

 Usage: dynamo.py [OPTIONS] COMMAND [ARGS]...                            
                                                                         
 Depth-Anything Dynamo CLI                                               
                                                                         
╭─ Commands ───────────────────────────────────────────────╮
│ export   Export Depth-Anything V2 using TS or Dynamo.    │
│ infer    Depth-Anything V2 inference using ONNXRuntime.  │
│          No dependency on PyTorch.                       │
╰──────────────────────────────────────────────────────────╯

If you would like to try out inference right away, you can download ONNX models that have already been exported here.

We observe the following average latencies using the CUDA Execution Provider:

Device	Encoder	Input Shape	Average Latency (ms)
RTX4080 12GB	ViT-S	`(1, 3, 518, 518)`	13.3
RTX4080 12GB	ViT-B	`(1, 3, 518, 518)`	29.3
RTX4080 12GB	ViT-L	`(1, 3, 518, 518)`	83.2

Relevant framework versions:

CUDA==12.1
cuDNN==8.9.2
onnxruntime-gpu==1.18.0
torch==2.3.1

Export Example

[!TIP] You can view the available options at any time by passing --help.

python dynamo.py export --encoder vitb --output weights/vitb.onnx --opset 17

For dynamic shapes:

python dynamo.py export --encoder vitb -b 0 -h 0 -w 0

Export via TorchDynamo Example

[!CAUTION] The TorchDynamo-based ONNX Exporter is a new beta feature that may undergo breaking changes in the future. Currently, only static shapes and opset version 18 are supported.

python dynamo.py export --encoder vitb --output weights/vitb.onnx --use-dynamo -h 1036 -w 1036

Inference Example

python dynamo.py infer weights/vitb.onnx -i assets/sacre_coeur1.jpg

This function serves as an implementation reference for performing inference with only ONNXRuntime and OpenCV as dependencies.

Legacy

<details> <summary> V1 </summary> ## 🔥 ONNX Export

Prior to exporting the ONNX models, please install the requirements.

To convert the Depth Anything models to ONNX, run export.py. The pretrained weights will be downloaded automatically.

<details> <summary>Export Example</summary> <pre> python export.py --model s </pre> </details>

If you would like to try out inference right away, you can download ONNX models that have already been exported here.

⚡ ONNX Inference

With ONNX models in hand, one can perform inference on Python using ONNX Runtime. See infer.py.

<details> <summary>Inference Example</summary> <pre> python infer.py --img assets/DSC_0410.JPG --model weights/depth_anything_vits14.onnx --viz </pre> </details>

⏱️ Inference Time Comparison

We report the inference time, or latency, of only the model; that is, the time taken for preprocessing, postprocessing, or copying data between the host & device is not measured. The average inference time is defined as the median over all samples in the MegaDepth test dataset. We use the data provided by LoFTR here - a total of 806 images.

Each image is resized such that its size is 518x518 before being fed into the model. The inference time is then measured for all model variants (S, B, L). See eval.py for the measurement code.

All experiments are conducted on an i9-12900HX CPU and RTX4080 12GB GPU with CUDA==11.8.1, torch==2.1.2, and onnxruntime==1.16.3.

Notes

Currently, the inference speed is bottlenecked by Conv operations.
ONNXRuntime performs slightly (20-25%) faster for the ViT-L model variant.

</details>

Credits

If you use any ideas from the papers or code in this repo, please consider citing the authors of Depth Anything, Depth Anything V2 and DINOv2. Lastly, if the ONNX versions helped you in any way, please also consider starring this repository.

@article{yang2024depth,
      title={Depth Anything V2}, 
      author={Lihe Yang and Bingyi Kang and Zilong Huang and Zhen Zhao and Xiaogang Xu and Jiashi Feng and Hengshuang Zhao},
      year={2024},
      eprint={2406.09414},
      archivePrefix={arXiv},
      primaryClass={id='cs.CV' full_name='Computer Vision and Pattern Recognition' is_active=True alt_name=None in_archive='cs' is_general=False description='Covers image processing, computer vision, pattern recognition, and scene understanding. Roughly includes material in ACM Subject Classes I.2.10, I.4, and I.5.'}
}

@article{depthanything,
      title={Depth Anything: Unleashing the Power of Large-Scale Unlabeled Data}, 
      author={Yang, Lihe and Kang, Bingyi and Huang, Zilong and Xu, Xiaogang and Feng, Jiashi and Zhao, Hengshuang},
      journal={arXiv:2401.10891},
      year={2024}
}

@misc{oquab2023dinov2,
  title={DINOv2: Learning Robust Visual Features without Supervision},
  author={Oquab, Maxime and Darcet, Timothée and Moutakanni, Theo and Vo, Huy V. and Szafraniec, Marc and Khalidov, Vasil and Fernandez, Pierre and Haziza, Daniel and Massa, Francisco and El-Nouby, Alaaeldin and Howes, Russell and Huang, Po-Yao and Xu, Hu and Sharma, Vasu and Li, Shang-Wen and Galuba, Wojciech and Rabbat, Mike and Assran, Mido and Ballas, Nicolas and Synnaeve, Gabriel and Misra, Ishan and Jegou, Herve and Mairal, Julien and Labatut, Patrick and Joulin, Armand and Bojanowski, Piotr},
  journal={arXiv:2304.07193},
  year={2023}
}