amael-apple
commited on
Commit
•
41f26b3
1
Parent(s):
19ad76e
Upload folder using huggingface_hub
Browse files- README.md +135 -3
- config.json +7 -0
- model.safetensors +3 -0
README.md
CHANGED
@@ -1,3 +1,135 @@
|
|
1 |
-
---
|
2 |
-
license: apple-ascl
|
3 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
---
|
2 |
+
license: apple-ascl
|
3 |
+
pipeline_tag: depth-estimation
|
4 |
+
tags:
|
5 |
+
- model_hub_mixin
|
6 |
+
- pytorch_model_hub_mixin
|
7 |
+
---
|
8 |
+
|
9 |
+
# Depth Pro: Sharp Monocular Metric Depth in Less Than a Second
|
10 |
+
|
11 |
+
![Depth Pro Demo Image](https://github.com/apple/ml-depth-pro/raw/main/data/depth-pro-teaser.jpg)
|
12 |
+
|
13 |
+
We present a foundation model for zero-shot metric monocular depth estimation. Our model, Depth Pro, synthesizes high-resolution depth maps with unparalleled sharpness and high-frequency details. The predictions are metric, with absolute scale, without relying on the availability of metadata such as camera intrinsics. And the model is fast, producing a 2.25-megapixel depth map in 0.3 seconds on a standard GPU. These characteristics are enabled by a number of technical contributions, including an efficient multi-scale vision transformer for dense prediction, a training protocol that combines real and synthetic datasets to achieve high metric accuracy alongside fine boundary tracing, dedicated evaluation metrics for boundary accuracy in estimated depth maps, and state-of-the-art focal length estimation from a single image.
|
14 |
+
|
15 |
+
Depth Pro was introduced in **[Depth Pro: Sharp Monocular Metric Depth in Less Than a Second](https://arxiv.org/abs/2410.02073)**, by *Aleksei Bochkovskii, Amaël Delaunoy, Hugo Germain, Marcel Santos, Yichao Zhou, Stephan R. Richter, and Vladlen Koltun*.
|
16 |
+
|
17 |
+
The checkpoint in this repository is a reference implementation, which has been re-trained. Its performance is close to the model reported in the paper but does not match it exactly.
|
18 |
+
|
19 |
+
## How to Use
|
20 |
+
|
21 |
+
Please, follow the steps in the [code repository](https://github.com/apple/ml-depth-pro) to set up your environment. Then you can:
|
22 |
+
|
23 |
+
### Running from Python
|
24 |
+
|
25 |
+
```python
|
26 |
+
from huggingface_hub import PyTorchModelHubMixin
|
27 |
+
from depth_pro import create_model_and_transforms, load_rgb
|
28 |
+
from depth_pro.depth_pro import (create_backbone_model, load_monodepth_weights,
|
29 |
+
DepthPro, DepthProEncoder, MultiresConvDecoder)
|
30 |
+
import depth_pro
|
31 |
+
from torchvision.transforms import Compose, Normalize
|
32 |
+
|
33 |
+
|
34 |
+
class DepthProWrapper(DepthPro, PyTorchModelHubMixin):
|
35 |
+
"""Depth Pro network."""
|
36 |
+
|
37 |
+
def __init__(
|
38 |
+
self,
|
39 |
+
patch_encoder_preset: str,
|
40 |
+
image_encoder_preset: str,
|
41 |
+
decoder_features: str,
|
42 |
+
fov_encoder_preset: str,
|
43 |
+
use_fov_head: bool = True,
|
44 |
+
**kwargs,
|
45 |
+
):
|
46 |
+
"""Initialize Depth Pro."""
|
47 |
+
|
48 |
+
patch_encoder, patch_encoder_config = create_backbone_model(
|
49 |
+
preset=patch_encoder_preset
|
50 |
+
)
|
51 |
+
image_encoder, _ = create_backbone_model(
|
52 |
+
preset=image_encoder_preset
|
53 |
+
)
|
54 |
+
|
55 |
+
fov_encoder = None
|
56 |
+
if use_fov_head and fov_encoder_preset is not None:
|
57 |
+
fov_encoder, _ = create_backbone_model(preset=fov_encoder_preset)
|
58 |
+
|
59 |
+
dims_encoder = patch_encoder_config.encoder_feature_dims
|
60 |
+
hook_block_ids = patch_encoder_config.encoder_feature_layer_ids
|
61 |
+
encoder = DepthProEncoder(
|
62 |
+
dims_encoder=dims_encoder,
|
63 |
+
patch_encoder=patch_encoder,
|
64 |
+
image_encoder=image_encoder,
|
65 |
+
hook_block_ids=hook_block_ids,
|
66 |
+
decoder_features=decoder_features,
|
67 |
+
)
|
68 |
+
decoder = MultiresConvDecoder(
|
69 |
+
dims_encoder=[encoder.dims_encoder[0]] + list(encoder.dims_encoder),
|
70 |
+
dim_decoder=decoder_features,
|
71 |
+
)
|
72 |
+
|
73 |
+
super().__init__(
|
74 |
+
encoder=encoder,
|
75 |
+
decoder=decoder,
|
76 |
+
last_dims=(32, 1),
|
77 |
+
use_fov_head=use_fov_head,
|
78 |
+
fov_encoder=fov_encoder,
|
79 |
+
)
|
80 |
+
|
81 |
+
|
82 |
+
# Load model and preprocessing transform
|
83 |
+
model = DepthProWrapper.from_pretrained("DepthPro-L")
|
84 |
+
transform = Compose(
|
85 |
+
[
|
86 |
+
ToTensor(),
|
87 |
+
Normalize([0.5, 0.5, 0.5], [0.5, 0.5, 0.5]),
|
88 |
+
]
|
89 |
+
)
|
90 |
+
|
91 |
+
|
92 |
+
model.eval()
|
93 |
+
|
94 |
+
# Load and preprocess an image.
|
95 |
+
image, _, f_px = depth_pro.load_rgb(image_path)
|
96 |
+
image = transform(image)
|
97 |
+
|
98 |
+
# Run inference.
|
99 |
+
prediction = model.infer(image, f_px=f_px)
|
100 |
+
depth = prediction["depth"] # Depth in [m].
|
101 |
+
focallength_px = prediction["focallength_px"] # Focal length in pixels.
|
102 |
+
```
|
103 |
+
|
104 |
+
### Evaluation (boundary metrics)
|
105 |
+
|
106 |
+
Boundary metrics are implemented in `eval/boundary_metrics.py` and can be used as follows:
|
107 |
+
|
108 |
+
```python
|
109 |
+
# for a depth-based dataset
|
110 |
+
boundary_f1 = SI_boundary_F1(predicted_depth, target_depth)
|
111 |
+
|
112 |
+
# for a mask-based dataset (image matting / segmentation)
|
113 |
+
boundary_recall = SI_boundary_Recall(predicted_depth, target_mask)
|
114 |
+
```
|
115 |
+
|
116 |
+
|
117 |
+
## Citation
|
118 |
+
|
119 |
+
If you find our work useful, please cite the following paper:
|
120 |
+
|
121 |
+
```bibtex
|
122 |
+
@article{Bochkovskii2024:arxiv,
|
123 |
+
author = {Aleksei Bochkovskii and Ama\"{e}l Delaunoy and Hugo Germain and Marcel Santos and
|
124 |
+
Yichao Zhou and Stephan R. Richter and Vladlen Koltun}
|
125 |
+
title = {Depth Pro: Sharp Monocular Metric Depth in Less Than a Second},
|
126 |
+
journal = {arXiv},
|
127 |
+
year = {2024},
|
128 |
+
}
|
129 |
+
```
|
130 |
+
|
131 |
+
## Acknowledgements
|
132 |
+
|
133 |
+
Our codebase is built using multiple opensource contributions, please see [Acknowledgements](https://github.com/apple/ml-depth-pro/blob/main/ACKNOWLEDGEMENTS.md) for more details.
|
134 |
+
|
135 |
+
Please check the paper for a complete list of references and datasets used in this work.
|
config.json
ADDED
@@ -0,0 +1,7 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"decoder_features": 256,
|
3 |
+
"fov_encoder_preset": "dinov2l16_384",
|
4 |
+
"image_encoder_preset": "dinov2l16_384",
|
5 |
+
"patch_encoder_preset": "dinov2l16_384",
|
6 |
+
"use_fov_head": true
|
7 |
+
}
|
model.safetensors
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:8cf414ab41135c007626ebde7013252279628de1de2bc9579cce5bc49127d33f
|
3 |
+
size 1904109940
|