README.md
| 1 | --- |
| 2 | library_name: transformers |
| 3 | license: mit |
| 4 | pipeline_tag: depth-estimation |
| 5 | arxiv: <2502.19204> |
| 6 | tags: |
| 7 | - distill-any-depth |
| 8 | - vision |
| 9 | --- |
| 10 | |
| 11 | # Distill Any Depth Large - Transformers Version |
| 12 | |
| 13 | ## Introduction |
| 14 | We present Distill-Any-Depth, a new SOTA monocular depth estimation model trained with our proposed knowledge distillation algorithms. It was introduced in the paper [Distill Any Depth: Distillation Creates a Stronger Monocular Depth Estimator](http://arxiv.org/abs/2502.19204). |
| 15 | |
| 16 | This model checkpoint is compatible with the transformers library. |
| 17 | |
| 18 | [Online demo](https://huggingface.co/spaces/xingyang1/Distill-Any-Depth). |
| 19 | |
| 20 | ### How to use |
| 21 | |
| 22 | Here is how to use this model to perform zero-shot depth estimation: |
| 23 | |
| 24 | ```python |
| 25 | from transformers import pipeline |
| 26 | from PIL import Image |
| 27 | import requests |
| 28 | # load pipe |
| 29 | pipe = pipeline(task="depth-estimation", model="xingyang1/Distill-Any-Depth-Large-hf") |
| 30 | # load image |
| 31 | url = 'http://images.cocodataset.org/val2017/000000039769.jpg' |
| 32 | image = Image.open(requests.get(url, stream=True).raw) |
| 33 | # inference |
| 34 | depth = pipe(image)["depth"] |
| 35 | ``` |
| 36 | |
| 37 | Alternatively, you can use the model and processor classes: |
| 38 | |
| 39 | ```python |
| 40 | from transformers import AutoImageProcessor, AutoModelForDepthEstimation |
| 41 | import torch |
| 42 | import numpy as np |
| 43 | from PIL import Image |
| 44 | import requests |
| 45 | |
| 46 | url = "http://images.cocodataset.org/val2017/000000039769.jpg" |
| 47 | image = Image.open(requests.get(url, stream=True).raw) |
| 48 | |
| 49 | image_processor = AutoImageProcessor.from_pretrained("xingyang1/Distill-Any-Depth-Large-hf") |
| 50 | model = AutoModelForDepthEstimation.from_pretrained("xingyang1/Distill-Any-Depth-Large-hf") |
| 51 | |
| 52 | # prepare image for the model |
| 53 | inputs = image_processor(images=image, return_tensors="pt") |
| 54 | |
| 55 | with torch.no_grad(): |
| 56 | outputs = model(**inputs) |
| 57 | |
| 58 | # interpolate to original size and visualize the prediction |
| 59 | post_processed_output = image_processor.post_process_depth_estimation( |
| 60 | outputs, |
| 61 | target_sizes=[(image.height, image.width)], |
| 62 | ) |
| 63 | |
| 64 | predicted_depth = post_processed_output[0]["predicted_depth"] |
| 65 | depth = (predicted_depth - predicted_depth.min()) / (predicted_depth.max() - predicted_depth.min()) |
| 66 | depth = depth.detach().cpu().numpy() * 255 |
| 67 | depth = Image.fromarray(depth.astype("uint8")) |
| 68 | ) |
| 69 | ``` |
| 70 | |
| 71 | |
| 72 | If you find this project useful, please consider citing: |
| 73 | |
| 74 | ```bibtex |
| 75 | @article{he2025distill, |
| 76 | title = {Distill Any Depth: Distillation Creates a Stronger Monocular Depth Estimator}, |
| 77 | author = {Xiankang He and Dongyan Guo and Hongji Li and Ruibo Li and Ying Cui and Chi Zhang}, |
| 78 | year = {2025}, |
| 79 | journal = {arXiv preprint arXiv: 2502.19204} |
| 80 | } |
| 81 | ``` |
| 82 | |
| 83 | ## Model Card Author |
| 84 | [Parteek Kamboj](https://huggingface.co/keetrap) |