README.md
7.7 KB · 166 lines · markdown Raw
1 ---
2 license: apache-2.0
3 tags:
4 - vision
5 - depth-estimation
6 widget:
7 - src: https://huggingface.co/datasets/mishig/sample_images/resolve/main/tiger.jpg
8 example_title: Tiger
9 - src: https://huggingface.co/datasets/mishig/sample_images/resolve/main/teapot.jpg
10 example_title: Teapot
11 - src: https://huggingface.co/datasets/mishig/sample_images/resolve/main/palace.jpg
12 example_title: Palace
13 model-index:
14 - name: dpt-hybrid-midas
15 results:
16 - task:
17 type: monocular-depth-estimation
18 name: Monocular Depth Estimation
19 dataset:
20 type: MIX-6
21 name: MIX-6
22 metrics:
23 - type: Zero-shot transfer
24 value: 11.06
25 name: Zero-shot transfer
26 config: Zero-shot transfer
27 verified: false
28
29 ---
30
31 ## Model Details: DPT-Hybrid (also known as MiDaS 3.0)
32
33 Dense Prediction Transformer (DPT) model trained on 1.4 million images for monocular depth estimation.
34 It was introduced in the paper [Vision Transformers for Dense Prediction](https://arxiv.org/abs/2103.13413) by Ranftl et al. (2021) and first released in [this repository](https://github.com/isl-org/DPT).
35 DPT uses the Vision Transformer (ViT) as backbone and adds a neck + head on top for monocular depth estimation.
36 ![model image](https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/dpt_architecture.jpg)
37
38 This repository hosts the "hybrid" version of the model as stated in the paper. DPT-Hybrid diverges from DPT by using [ViT-hybrid](https://huggingface.co/google/vit-hybrid-base-bit-384) as a backbone and taking some activations from the backbone.
39
40 The model card has been written in combination by the Hugging Face team and Intel.
41
42 | Model Detail | Description |
43 | ----------- | ----------- |
44 | Model Authors - Company | Intel |
45 | Date | December 22, 2022 |
46 | Version | 1 |
47 | Type | Computer Vision - Monocular Depth Estimation |
48 | Paper or Other Resources | [Vision Transformers for Dense Prediction](https://arxiv.org/abs/2103.13413) and [GitHub Repo](https://github.com/isl-org/DPT) |
49 | License | Apache 2.0 |
50 | Questions or Comments | [Community Tab](https://huggingface.co/Intel/dpt-hybrid-midas/discussions) and [Intel Developers Discord](https://discord.gg/rv2Gp55UJQ)|
51
52 | Intended Use | Description |
53 | ----------- | ----------- |
54 | Primary intended uses | You can use the raw model for zero-shot monocular depth estimation. See the [model hub](https://huggingface.co/models?search=dpt) to look for fine-tuned versions on a task that interests you. |
55 | Primary intended users | Anyone doing monocular depth estimation |
56 | Out-of-scope uses | This model in most cases will need to be fine-tuned for your particular task. The model should not be used to intentionally create hostile or alienating environments for people.|
57
58 ### How to use
59
60 Here is how to use this model for zero-shot depth estimation on an image:
61
62 ```python
63 from PIL import Image
64 import numpy as np
65 import requests
66 import torch
67
68 from transformers import DPTImageProcessor, DPTForDepthEstimation
69
70 image_processor = DPTImageProcessor.from_pretrained("Intel/dpt-hybrid-midas")
71 model = DPTForDepthEstimation.from_pretrained("Intel/dpt-hybrid-midas", low_cpu_mem_usage=True)
72
73 url = "http://images.cocodataset.org/val2017/000000039769.jpg"
74 image = Image.open(requests.get(url, stream=True).raw)
75
76 # prepare image for the model
77 inputs = image_processor(images=image, return_tensors="pt")
78
79 with torch.no_grad():
80 outputs = model(**inputs)
81 predicted_depth = outputs.predicted_depth
82
83 # interpolate to original size
84 prediction = torch.nn.functional.interpolate(
85 predicted_depth.unsqueeze(1),
86 size=image.size[::-1],
87 mode="bicubic",
88 align_corners=False,
89 )
90
91 # visualize the prediction
92 output = prediction.squeeze().cpu().numpy()
93 formatted = (output * 255 / np.max(output)).astype("uint8")
94 depth = Image.fromarray(formatted)
95 depth.show()
96 ```
97
98 For more code examples, we refer to the [documentation](https://huggingface.co/docs/transformers/master/en/model_doc/dpt).
99
100 | Factors | Description |
101 | ----------- | ----------- |
102 | Groups | Multiple datasets compiled together |
103 | Instrumentation | - |
104 | Environment | Inference completed on Intel Xeon Platinum 8280 CPU @ 2.70GHz with 8 physical cores and an NVIDIA RTX 2080 GPU. |
105 | Card Prompts | Model deployment on alternate hardware and software will change model performance |
106
107 | Metrics | Description |
108 | ----------- | ----------- |
109 | Model performance measures | Zero-shot Transfer |
110 | Decision thresholds | - |
111 | Approaches to uncertainty and variability | - |
112
113 | Training and Evaluation Data | Description |
114 | ----------- | ----------- |
115 | Datasets | The dataset is called MIX 6, and contains around 1.4M images. The model was initialized with ImageNet-pretrained weights.|
116 | Motivation | To build a robust monocular depth prediction network |
117 | Preprocessing | "We resize the image such that the longer side is 384 pixels and train on random square crops of size 384. ... We perform random horizontal flips for data augmentation." See [Ranftl et al. (2021)](https://arxiv.org/abs/2103.13413) for more details. |
118
119 ## Quantitative Analyses
120 | Model | Training set | DIW WHDR | ETH3D AbsRel | Sintel AbsRel | KITTI δ>1.25 | NYU δ>1.25 | TUM δ>1.25 |
121 | --- | --- | --- | --- | --- | --- | --- | --- |
122 | DPT - Large | MIX 6 | 10.82 (-13.2%) | 0.089 (-31.2%) | 0.270 (-17.5%) | 8.46 (-64.6%) | 8.32 (-12.9%) | 9.97 (-30.3%) |
123 | DPT - Hybrid | MIX 6 | 11.06 (-11.2%) | 0.093 (-27.6%) | 0.274 (-16.2%) | 11.56 (-51.6%) | 8.69 (-9.0%) | 10.89 (-23.2%) |
124 | MiDaS | MIX 6 | 12.95 (+3.9%) | 0.116 (-10.5%) | 0.329 (+0.5%) | 16.08 (-32.7%) | 8.71 (-8.8%) | 12.51 (-12.5%)
125 | MiDaS [30] | MIX 5 | 12.46 | 0.129 | 0.327 | 23.90 | 9.55 | 14.29 |
126 | Li [22] | MD [22] | 23.15 | 0.181 | 0.385 | 36.29 | 27.52 | 29.54 |
127 | Li [21] | MC [21] | 26.52 | 0.183 | 0.405 | 47.94 | 18.57 | 17.71 |
128 | Wang [40] | WS [40] | 19.09 | 0.205 | 0.390 | 31.92 | 29.57 | 20.18 |
129 | Xian [45] | RW [45] | 14.59 | 0.186 | 0.422 | 34.08 | 27.00 | 25.02 |
130 | Casser [5] | CS [8] | 32.80 | 0.235 | 0.422 | 21.15 | 39.58 | 37.18 |
131
132 Table 1. Comparison to the state of the art on monocular depth estimation. We evaluate zero-shot cross-dataset transfer according to the
133 protocol defined in [30]. Relative performance is computed with respect to the original MiDaS model [30]. Lower is better for all metrics. ([Ranftl et al., 2021](https://arxiv.org/abs/2103.13413))
134
135
136 | Ethical Considerations | Description |
137 | ----------- | ----------- |
138 | Data | The training data come from multiple image datasets compiled together. |
139 | Human life | The model is not intended to inform decisions central to human life or flourishing. It is an aggregated set of monocular depth image datasets. |
140 | Mitigations | No additional risk mitigation strategies were considered during model development. |
141 | Risks and harms | The extent of the risks involved by using the model remain unknown. |
142 | Use cases | - |
143
144 | Caveats and Recommendations |
145 | ----------- |
146 | Users (both direct and downstream) should be made aware of the risks, biases and limitations of the model. There are no additional caveats or recommendations for this model. |
147
148 ### BibTeX entry and citation info
149
150 ```bibtex
151 @article{DBLP:journals/corr/abs-2103-13413,
152 author = {Ren{\'{e}} Ranftl and
153 Alexey Bochkovskiy and
154 Vladlen Koltun},
155 title = {Vision Transformers for Dense Prediction},
156 journal = {CoRR},
157 volume = {abs/2103.13413},
158 year = {2021},
159 url = {https://arxiv.org/abs/2103.13413},
160 eprinttype = {arXiv},
161 eprint = {2103.13413},
162 timestamp = {Wed, 07 Apr 2021 15:31:46 +0200},
163 biburl = {https://dblp.org/rec/journals/corr/abs-2103-13413.bib},
164 bibsource = {dblp computer science bibliography, https://dblp.org}
165 }
166 ```