Spaces:
Running
on
Zero
Running
on
Zero
update the pipeline to the latest one from diffusers
Browse filesupdate gradio to 4.21.0
add cvpr acceptance note
support gradio spaces for zero gpu
fix 3d printable artefact to have horizontal orientation
change minimum denoising steps to 10
add a note with the pointer to Marigold-LCM
- .gitattributes +3 -2
- README.md +8 -10
- app.py +100 -46
- extrude.py +33 -1
- files/bee.jpg +0 -0
- files/cat.jpg +0 -0
- files/swings.jpg +0 -0
- marigold_depth_estimation.py +632 -0
- requirements.txt +4 -4
.gitattributes
CHANGED
@@ -33,5 +33,6 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
|
|
33 |
*.zip filter=lfs diff=lfs merge=lfs -text
|
34 |
*.zst filter=lfs diff=lfs merge=lfs -text
|
35 |
*tfevents* filter=lfs diff=lfs merge=lfs -text
|
36 |
-
|
37 |
-
|
|
|
|
33 |
*.zip filter=lfs diff=lfs merge=lfs -text
|
34 |
*.zst filter=lfs diff=lfs merge=lfs -text
|
35 |
*tfevents* filter=lfs diff=lfs merge=lfs -text
|
36 |
+
*.npy filter=lfs diff=lfs merge=lfs -text
|
37 |
+
*.jpg filter=lfs diff=lfs merge=lfs -text
|
38 |
+
*.png filter=lfs diff=lfs merge=lfs -text
|
README.md
CHANGED
@@ -4,23 +4,21 @@ emoji: 🏵️
|
|
4 |
colorFrom: blue
|
5 |
colorTo: red
|
6 |
sdk: gradio
|
7 |
-
sdk_version: 4.
|
8 |
app_file: app.py
|
9 |
pinned: true
|
10 |
license: cc-by-sa-4.0
|
11 |
models:
|
12 |
-
-
|
13 |
---
|
14 |
|
15 |
-
This is a demo of the monocular depth estimation pipeline, described in the paper titled ["Repurposing Diffusion-Based Image Generators for Monocular Depth Estimation"](https://arxiv.org/abs/2312.02145)
|
16 |
|
17 |
```
|
18 |
-
@
|
19 |
-
|
20 |
-
|
21 |
-
|
22 |
-
|
23 |
-
archivePrefix={arXiv},
|
24 |
-
primaryClass={cs.CV}
|
25 |
}
|
26 |
```
|
|
|
4 |
colorFrom: blue
|
5 |
colorTo: red
|
6 |
sdk: gradio
|
7 |
+
sdk_version: 4.21.0
|
8 |
app_file: app.py
|
9 |
pinned: true
|
10 |
license: cc-by-sa-4.0
|
11 |
models:
|
12 |
+
- prs-eth/marigold-v1-0
|
13 |
---
|
14 |
|
15 |
+
This is a demo of the monocular depth estimation pipeline, described in the CVPR 2024 paper titled ["Repurposing Diffusion-Based Image Generators for Monocular Depth Estimation"](https://arxiv.org/abs/2312.02145)
|
16 |
|
17 |
```
|
18 |
+
@InProceedings{ke2023repurposing,
|
19 |
+
title={Repurposing Diffusion-Based Image Generators for Monocular Depth Estimation},
|
20 |
+
author={Bingxin Ke and Anton Obukhov and Shengyu Huang and Nando Metzger and Rodrigo Caye Daudt and Konrad Schindler},
|
21 |
+
booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
|
22 |
+
year={2024}
|
|
|
|
|
23 |
}
|
24 |
```
|
app.py
CHANGED
@@ -1,17 +1,37 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
import functools
|
2 |
import os
|
3 |
-
import shutil
|
4 |
-
import sys
|
5 |
|
6 |
-
import
|
7 |
import gradio as gr
|
8 |
import numpy as np
|
9 |
import torch as torch
|
10 |
from PIL import Image
|
11 |
|
12 |
from gradio_imageslider import ImageSlider
|
|
|
13 |
|
14 |
from extrude import extrude_depth_3d
|
|
|
15 |
|
16 |
|
17 |
def process(
|
@@ -82,12 +102,21 @@ def process_3d(
|
|
82 |
frame_far,
|
83 |
):
|
84 |
if input_image is None or len(files) < 1:
|
85 |
-
raise gr.Error(
|
|
|
|
|
86 |
|
87 |
if plane_near >= plane_far:
|
88 |
raise gr.Error("NEAR plane must have a value smaller than the FAR plane")
|
89 |
|
90 |
-
def _process_3d(
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
91 |
image_rgb = input_image
|
92 |
image_depth = files[0]
|
93 |
|
@@ -105,14 +134,18 @@ def process_3d(
|
|
105 |
image_rgb_content.resize((image_new_w, image_new_h), Image.LANCZOS).save(
|
106 |
image_rgb_new
|
107 |
)
|
108 |
-
Image.open(image_depth).resize((image_new_w, image_new_h), Image.
|
109 |
image_depth_new
|
110 |
)
|
111 |
|
112 |
path_glb, path_stl = extrude_depth_3d(
|
113 |
image_rgb_new,
|
114 |
image_depth_new,
|
115 |
-
output_model_scale=
|
|
|
|
|
|
|
|
|
116 |
filter_size=filter_size,
|
117 |
coef_near=plane_near,
|
118 |
coef_far=plane_far,
|
@@ -122,24 +155,27 @@ def process_3d(
|
|
122 |
f_back=frame_far / 100,
|
123 |
vertex_colors=vertex_colors,
|
124 |
scene_lights=scene_lights,
|
|
|
125 |
)
|
126 |
|
127 |
return path_glb, path_stl
|
128 |
|
129 |
-
path_viewer_glb, _ = _process_3d(
|
130 |
-
|
131 |
-
|
132 |
-
|
133 |
-
|
134 |
-
|
135 |
-
|
136 |
-
|
|
|
|
|
137 |
|
138 |
return path_viewer_glb, [path_files_glb, path_files_stl]
|
139 |
|
140 |
|
141 |
def run_demo_server(pipe):
|
142 |
-
process_pipe = functools.partial(process, pipe)
|
143 |
os.environ["GRADIO_ALLOW_FLAGGING"] = "never"
|
144 |
|
145 |
with gr.Blocks(
|
@@ -156,11 +192,24 @@ def run_demo_server(pipe):
|
|
156 |
.viewport {
|
157 |
aspect-ratio: 4/3;
|
158 |
}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
159 |
""",
|
160 |
) as demo:
|
161 |
gr.Markdown(
|
162 |
"""
|
163 |
-
|
|
|
164 |
<p align="center">
|
165 |
<a title="Website" href="https://marigoldmonodepth.github.io/" target="_blank" rel="noopener noreferrer" style="display: inline-block;">
|
166 |
<img src="https://www.obukhov.ai/img/badges/badge-website.svg">
|
@@ -175,12 +224,15 @@ def run_demo_server(pipe):
|
|
175 |
<img src="https://www.obukhov.ai/img/badges/badge-social.svg" alt="social">
|
176 |
</a>
|
177 |
</p>
|
178 |
-
|
179 |
-
|
180 |
-
|
181 |
-
|
182 |
-
|
183 |
-
|
|
|
|
|
|
|
184 |
"""
|
185 |
)
|
186 |
|
@@ -200,7 +252,7 @@ def run_demo_server(pipe):
|
|
200 |
)
|
201 |
denoise_steps = gr.Slider(
|
202 |
label="Number of denoising steps",
|
203 |
-
minimum=
|
204 |
maximum=20,
|
205 |
step=1,
|
206 |
value=10,
|
@@ -356,8 +408,17 @@ def run_demo_server(pipe):
|
|
356 |
)
|
357 |
|
358 |
blocks_settings_depth = [ensemble_size, denoise_steps, processing_res]
|
359 |
-
blocks_settings_3d = [
|
360 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
361 |
blocks_settings = blocks_settings_depth + blocks_settings_3d
|
362 |
map_id_to_default = {b._id: b.value for b in blocks_settings}
|
363 |
|
@@ -470,14 +531,21 @@ def run_demo_server(pipe):
|
|
470 |
gr.Button(interactive=True),
|
471 |
gr.Button(interactive=True),
|
472 |
gr.Image(value=None, interactive=True),
|
473 |
-
None,
|
|
|
|
|
|
|
|
|
|
|
|
|
474 |
]
|
475 |
return out
|
476 |
|
477 |
clear_btn.click(
|
478 |
fn=clear_fn,
|
479 |
inputs=[],
|
480 |
-
outputs=blocks_settings
|
|
|
481 |
submit_btn,
|
482 |
submit_3d,
|
483 |
input_image,
|
@@ -532,37 +600,23 @@ def run_demo_server(pipe):
|
|
532 |
)
|
533 |
|
534 |
|
535 |
-
def prefetch_hf_cache(pipe):
|
536 |
-
process(pipe, "files/bee.jpg", 1, 1, 64)
|
537 |
-
shutil.rmtree("files/bee_output")
|
538 |
-
|
539 |
-
|
540 |
def main():
|
541 |
-
|
542 |
-
REPO_HASH = "02cdfa52"
|
543 |
-
REPO_DIR = "Marigold"
|
544 |
-
CHECKPOINT = "Bingxin/Marigold"
|
545 |
|
546 |
-
if os.
|
547 |
-
|
548 |
-
repo = git.Repo.clone_from(REPO_URL, REPO_DIR)
|
549 |
-
repo.git.checkout(REPO_HASH)
|
550 |
-
|
551 |
-
sys.path.append(os.path.join(os.getcwd(), REPO_DIR))
|
552 |
-
|
553 |
-
from marigold import MarigoldPipeline
|
554 |
|
555 |
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
556 |
|
557 |
pipe = MarigoldPipeline.from_pretrained(CHECKPOINT)
|
558 |
try:
|
559 |
import xformers
|
|
|
560 |
pipe.enable_xformers_memory_efficient_attention()
|
561 |
except:
|
562 |
pass # run without xformers
|
563 |
|
564 |
pipe = pipe.to(device)
|
565 |
-
prefetch_hf_cache(pipe)
|
566 |
run_demo_server(pipe)
|
567 |
|
568 |
|
|
|
1 |
+
# Copyright 2024 Anton Obukhov, ETH Zurich. All rights reserved.
|
2 |
+
#
|
3 |
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
4 |
+
# you may not use this file except in compliance with the License.
|
5 |
+
# You may obtain a copy of the License at
|
6 |
+
#
|
7 |
+
# http://www.apache.org/licenses/LICENSE-2.0
|
8 |
+
#
|
9 |
+
# Unless required by applicable law or agreed to in writing, software
|
10 |
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
11 |
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
12 |
+
# See the License for the specific language governing permissions and
|
13 |
+
# limitations under the License.
|
14 |
+
# --------------------------------------------------------------------------
|
15 |
+
# If you find this code useful, we kindly ask you to cite our paper in your work.
|
16 |
+
# Please find bibtex at: https://github.com/prs-eth/Marigold#-citation
|
17 |
+
# More information about the method can be found at https://marigoldmonodepth.github.io
|
18 |
+
# --------------------------------------------------------------------------
|
19 |
+
|
20 |
+
|
21 |
import functools
|
22 |
import os
|
|
|
|
|
23 |
|
24 |
+
import spaces
|
25 |
import gradio as gr
|
26 |
import numpy as np
|
27 |
import torch as torch
|
28 |
from PIL import Image
|
29 |
|
30 |
from gradio_imageslider import ImageSlider
|
31 |
+
from huggingface_hub import login
|
32 |
|
33 |
from extrude import extrude_depth_3d
|
34 |
+
from marigold_depth_estimation import MarigoldPipeline
|
35 |
|
36 |
|
37 |
def process(
|
|
|
102 |
frame_far,
|
103 |
):
|
104 |
if input_image is None or len(files) < 1:
|
105 |
+
raise gr.Error(
|
106 |
+
"Please upload an image (or use examples) and compute depth first"
|
107 |
+
)
|
108 |
|
109 |
if plane_near >= plane_far:
|
110 |
raise gr.Error("NEAR plane must have a value smaller than the FAR plane")
|
111 |
|
112 |
+
def _process_3d(
|
113 |
+
size_longest_px,
|
114 |
+
filter_size,
|
115 |
+
vertex_colors,
|
116 |
+
scene_lights,
|
117 |
+
output_model_scale=None,
|
118 |
+
prepare_for_3d_printing=False,
|
119 |
+
):
|
120 |
image_rgb = input_image
|
121 |
image_depth = files[0]
|
122 |
|
|
|
134 |
image_rgb_content.resize((image_new_w, image_new_h), Image.LANCZOS).save(
|
135 |
image_rgb_new
|
136 |
)
|
137 |
+
Image.open(image_depth).resize((image_new_w, image_new_h), Image.BILINEAR).save(
|
138 |
image_depth_new
|
139 |
)
|
140 |
|
141 |
path_glb, path_stl = extrude_depth_3d(
|
142 |
image_rgb_new,
|
143 |
image_depth_new,
|
144 |
+
output_model_scale=(
|
145 |
+
size_longest_cm * 10
|
146 |
+
if output_model_scale is None
|
147 |
+
else output_model_scale
|
148 |
+
),
|
149 |
filter_size=filter_size,
|
150 |
coef_near=plane_near,
|
151 |
coef_far=plane_far,
|
|
|
155 |
f_back=frame_far / 100,
|
156 |
vertex_colors=vertex_colors,
|
157 |
scene_lights=scene_lights,
|
158 |
+
prepare_for_3d_printing=prepare_for_3d_printing,
|
159 |
)
|
160 |
|
161 |
return path_glb, path_stl
|
162 |
|
163 |
+
path_viewer_glb, _ = _process_3d(
|
164 |
+
256, filter_size, vertex_colors=False, scene_lights=True, output_model_scale=1
|
165 |
+
)
|
166 |
+
path_files_glb, path_files_stl = _process_3d(
|
167 |
+
size_longest_px,
|
168 |
+
filter_size,
|
169 |
+
vertex_colors=True,
|
170 |
+
scene_lights=False,
|
171 |
+
prepare_for_3d_printing=True,
|
172 |
+
)
|
173 |
|
174 |
return path_viewer_glb, [path_files_glb, path_files_stl]
|
175 |
|
176 |
|
177 |
def run_demo_server(pipe):
|
178 |
+
process_pipe = spaces.GPU(functools.partial(process, pipe), duration=120)
|
179 |
os.environ["GRADIO_ALLOW_FLAGGING"] = "never"
|
180 |
|
181 |
with gr.Blocks(
|
|
|
192 |
.viewport {
|
193 |
aspect-ratio: 4/3;
|
194 |
}
|
195 |
+
h1 {
|
196 |
+
text-align: center;
|
197 |
+
display: block;
|
198 |
+
}
|
199 |
+
h2 {
|
200 |
+
text-align: center;
|
201 |
+
display: block;
|
202 |
+
}
|
203 |
+
h3 {
|
204 |
+
text-align: center;
|
205 |
+
display: block;
|
206 |
+
}
|
207 |
""",
|
208 |
) as demo:
|
209 |
gr.Markdown(
|
210 |
"""
|
211 |
+
# Marigold Depth Estimation
|
212 |
+
|
213 |
<p align="center">
|
214 |
<a title="Website" href="https://marigoldmonodepth.github.io/" target="_blank" rel="noopener noreferrer" style="display: inline-block;">
|
215 |
<img src="https://www.obukhov.ai/img/badges/badge-website.svg">
|
|
|
224 |
<img src="https://www.obukhov.ai/img/badges/badge-social.svg" alt="social">
|
225 |
</a>
|
226 |
</p>
|
227 |
+
|
228 |
+
Marigold is the state-of-the-art depth estimator for images in the wild.
|
229 |
+
Upload your image into the <b>first</b> pane, or click any of the <b>examples</b> below.
|
230 |
+
The result will be computed and appear in the <b>second</b> pane.
|
231 |
+
Scroll down to use the computed depth map for creating a 3D printable asset.
|
232 |
+
|
233 |
+
<a href="https://huggingface.co/spaces/prs-eth/marigold-lcm" style="color: crimson;">
|
234 |
+
<h3 style="color: crimson;">Check out Marigold-LCM — a FAST version of this demo!<h3>
|
235 |
+
</a>
|
236 |
"""
|
237 |
)
|
238 |
|
|
|
252 |
)
|
253 |
denoise_steps = gr.Slider(
|
254 |
label="Number of denoising steps",
|
255 |
+
minimum=10,
|
256 |
maximum=20,
|
257 |
step=1,
|
258 |
value=10,
|
|
|
408 |
)
|
409 |
|
410 |
blocks_settings_depth = [ensemble_size, denoise_steps, processing_res]
|
411 |
+
blocks_settings_3d = [
|
412 |
+
plane_near,
|
413 |
+
plane_far,
|
414 |
+
embossing,
|
415 |
+
size_longest_px,
|
416 |
+
size_longest_cm,
|
417 |
+
filter_size,
|
418 |
+
frame_thickness,
|
419 |
+
frame_near,
|
420 |
+
frame_far,
|
421 |
+
]
|
422 |
blocks_settings = blocks_settings_depth + blocks_settings_3d
|
423 |
map_id_to_default = {b._id: b.value for b in blocks_settings}
|
424 |
|
|
|
531 |
gr.Button(interactive=True),
|
532 |
gr.Button(interactive=True),
|
533 |
gr.Image(value=None, interactive=True),
|
534 |
+
None,
|
535 |
+
None,
|
536 |
+
None,
|
537 |
+
None,
|
538 |
+
None,
|
539 |
+
None,
|
540 |
+
None,
|
541 |
]
|
542 |
return out
|
543 |
|
544 |
clear_btn.click(
|
545 |
fn=clear_fn,
|
546 |
inputs=[],
|
547 |
+
outputs=blocks_settings
|
548 |
+
+ [
|
549 |
submit_btn,
|
550 |
submit_3d,
|
551 |
input_image,
|
|
|
600 |
)
|
601 |
|
602 |
|
|
|
|
|
|
|
|
|
|
|
603 |
def main():
|
604 |
+
CHECKPOINT = "prs-eth/marigold-v1-0"
|
|
|
|
|
|
|
605 |
|
606 |
+
if "HF_TOKEN_LOGIN" in os.environ:
|
607 |
+
login(token=os.environ["HF_TOKEN_LOGIN"])
|
|
|
|
|
|
|
|
|
|
|
|
|
608 |
|
609 |
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
610 |
|
611 |
pipe = MarigoldPipeline.from_pretrained(CHECKPOINT)
|
612 |
try:
|
613 |
import xformers
|
614 |
+
|
615 |
pipe.enable_xformers_memory_efficient_attention()
|
616 |
except:
|
617 |
pass # run without xformers
|
618 |
|
619 |
pipe = pipe.to(device)
|
|
|
620 |
run_demo_server(pipe)
|
621 |
|
622 |
|
extrude.py
CHANGED
@@ -1,3 +1,23 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
import math
|
2 |
import os
|
3 |
|
@@ -53,7 +73,12 @@ def glb_add_lights(path_input, path_output):
|
|
53 |
angle = i * angle_step
|
54 |
|
55 |
pos_rot = [0.0, 0.0, math.sin(angle / 2), math.cos(angle / 2)]
|
56 |
-
elev_rot = [
|
|
|
|
|
|
|
|
|
|
|
57 |
rotation = quaternion_multiply(pos_rot, elev_rot)
|
58 |
|
59 |
node = {
|
@@ -88,6 +113,7 @@ def extrude_depth_3d(
|
|
88 |
f_back=0.01,
|
89 |
vertex_colors=True,
|
90 |
scene_lights=True,
|
|
|
91 |
):
|
92 |
f_far_inner = -emboss
|
93 |
f_far_outer = f_far_inner - f_back
|
@@ -309,6 +335,12 @@ def extrude_depth_3d(
|
|
309 |
scaling_factor = output_model_scale / current_max_dimension
|
310 |
mesh.apply_scale(scaling_factor)
|
311 |
|
|
|
|
|
|
|
|
|
|
|
|
|
312 |
path_out_base = os.path.splitext(path_depth)[0].replace("_16bit", "")
|
313 |
path_out_glb = path_out_base + ".glb"
|
314 |
path_out_stl = path_out_base + ".stl"
|
|
|
1 |
+
# Copyright 2024 Anton Obukhov, ETH Zurich. All rights reserved.
|
2 |
+
#
|
3 |
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
4 |
+
# you may not use this file except in compliance with the License.
|
5 |
+
# You may obtain a copy of the License at
|
6 |
+
#
|
7 |
+
# http://www.apache.org/licenses/LICENSE-2.0
|
8 |
+
#
|
9 |
+
# Unless required by applicable law or agreed to in writing, software
|
10 |
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
11 |
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
12 |
+
# See the License for the specific language governing permissions and
|
13 |
+
# limitations under the License.
|
14 |
+
# --------------------------------------------------------------------------
|
15 |
+
# If you find this code useful, we kindly ask you to cite our paper in your work.
|
16 |
+
# Please find bibtex at: https://github.com/prs-eth/Marigold#-citation
|
17 |
+
# More information about the method can be found at https://marigoldmonodepth.github.io
|
18 |
+
# --------------------------------------------------------------------------
|
19 |
+
|
20 |
+
|
21 |
import math
|
22 |
import os
|
23 |
|
|
|
73 |
angle = i * angle_step
|
74 |
|
75 |
pos_rot = [0.0, 0.0, math.sin(angle / 2), math.cos(angle / 2)]
|
76 |
+
elev_rot = [
|
77 |
+
math.sin(elevation_angle / 2),
|
78 |
+
0.0,
|
79 |
+
0.0,
|
80 |
+
math.cos(elevation_angle / 2),
|
81 |
+
]
|
82 |
rotation = quaternion_multiply(pos_rot, elev_rot)
|
83 |
|
84 |
node = {
|
|
|
113 |
f_back=0.01,
|
114 |
vertex_colors=True,
|
115 |
scene_lights=True,
|
116 |
+
prepare_for_3d_printing=False,
|
117 |
):
|
118 |
f_far_inner = -emboss
|
119 |
f_far_outer = f_far_inner - f_back
|
|
|
335 |
scaling_factor = output_model_scale / current_max_dimension
|
336 |
mesh.apply_scale(scaling_factor)
|
337 |
|
338 |
+
if prepare_for_3d_printing:
|
339 |
+
rotation_mat = trimesh.transformations.rotation_matrix(
|
340 |
+
np.radians(90), [-1, 0, 0]
|
341 |
+
)
|
342 |
+
mesh.apply_transform(rotation_mat)
|
343 |
+
|
344 |
path_out_base = os.path.splitext(path_depth)[0].replace("_16bit", "")
|
345 |
path_out_glb = path_out_base + ".glb"
|
346 |
path_out_stl = path_out_base + ".stl"
|
files/bee.jpg
CHANGED
Git LFS Details
|
files/cat.jpg
CHANGED
Git LFS Details
|
files/swings.jpg
CHANGED
Git LFS Details
|
marigold_depth_estimation.py
ADDED
@@ -0,0 +1,632 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Copyright 2024 Bingxin Ke, ETH Zurich and The HuggingFace Team. All rights reserved.
|
2 |
+
#
|
3 |
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
4 |
+
# you may not use this file except in compliance with the License.
|
5 |
+
# You may obtain a copy of the License at
|
6 |
+
#
|
7 |
+
# http://www.apache.org/licenses/LICENSE-2.0
|
8 |
+
#
|
9 |
+
# Unless required by applicable law or agreed to in writing, software
|
10 |
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
11 |
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
12 |
+
# See the License for the specific language governing permissions and
|
13 |
+
# limitations under the License.
|
14 |
+
# --------------------------------------------------------------------------
|
15 |
+
# If you find this code useful, we kindly ask you to cite our paper in your work.
|
16 |
+
# Please find bibtex at: https://github.com/prs-eth/Marigold#-citation
|
17 |
+
# More information about the method can be found at https://marigoldmonodepth.github.io
|
18 |
+
# --------------------------------------------------------------------------
|
19 |
+
|
20 |
+
|
21 |
+
import math
|
22 |
+
from typing import Dict, Union
|
23 |
+
|
24 |
+
import matplotlib
|
25 |
+
import numpy as np
|
26 |
+
import torch
|
27 |
+
from PIL import Image
|
28 |
+
from scipy.optimize import minimize
|
29 |
+
from torch.utils.data import DataLoader, TensorDataset
|
30 |
+
from tqdm.auto import tqdm
|
31 |
+
from transformers import CLIPTextModel, CLIPTokenizer
|
32 |
+
|
33 |
+
from diffusers import (
|
34 |
+
AutoencoderKL,
|
35 |
+
DDIMScheduler,
|
36 |
+
DiffusionPipeline,
|
37 |
+
UNet2DConditionModel,
|
38 |
+
)
|
39 |
+
from diffusers.utils import BaseOutput, check_min_version
|
40 |
+
|
41 |
+
|
42 |
+
# Will error if the minimal version of diffusers is not installed. Remove at your own risks.
|
43 |
+
check_min_version("0.27.0.dev0")
|
44 |
+
|
45 |
+
|
46 |
+
class MarigoldDepthOutput(BaseOutput):
|
47 |
+
"""
|
48 |
+
Output class for Marigold monocular depth prediction pipeline.
|
49 |
+
|
50 |
+
Args:
|
51 |
+
depth_np (`np.ndarray`):
|
52 |
+
Predicted depth map, with depth values in the range of [0, 1].
|
53 |
+
depth_colored (`None` or `PIL.Image.Image`):
|
54 |
+
Colorized depth map, with the shape of [3, H, W] and values in [0, 1].
|
55 |
+
uncertainty (`None` or `np.ndarray`):
|
56 |
+
Uncalibrated uncertainty(MAD, median absolute deviation) coming from ensembling.
|
57 |
+
"""
|
58 |
+
|
59 |
+
depth_np: np.ndarray
|
60 |
+
depth_colored: Union[None, Image.Image]
|
61 |
+
uncertainty: Union[None, np.ndarray]
|
62 |
+
|
63 |
+
|
64 |
+
class MarigoldPipeline(DiffusionPipeline):
|
65 |
+
"""
|
66 |
+
Pipeline for monocular depth estimation using Marigold: https://marigoldmonodepth.github.io.
|
67 |
+
|
68 |
+
This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the
|
69 |
+
library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.)
|
70 |
+
|
71 |
+
Args:
|
72 |
+
unet (`UNet2DConditionModel`):
|
73 |
+
Conditional U-Net to denoise the depth latent, conditioned on image latent.
|
74 |
+
vae (`AutoencoderKL`):
|
75 |
+
Variational Auto-Encoder (VAE) Model to encode and decode images and depth maps
|
76 |
+
to and from latent representations.
|
77 |
+
scheduler (`DDIMScheduler`):
|
78 |
+
A scheduler to be used in combination with `unet` to denoise the encoded image latents.
|
79 |
+
text_encoder (`CLIPTextModel`):
|
80 |
+
Text-encoder, for empty text embedding.
|
81 |
+
tokenizer (`CLIPTokenizer`):
|
82 |
+
CLIP tokenizer.
|
83 |
+
"""
|
84 |
+
|
85 |
+
rgb_latent_scale_factor = 0.18215
|
86 |
+
depth_latent_scale_factor = 0.18215
|
87 |
+
|
88 |
+
def __init__(
|
89 |
+
self,
|
90 |
+
unet: UNet2DConditionModel,
|
91 |
+
vae: AutoencoderKL,
|
92 |
+
scheduler: DDIMScheduler,
|
93 |
+
text_encoder: CLIPTextModel,
|
94 |
+
tokenizer: CLIPTokenizer,
|
95 |
+
):
|
96 |
+
super().__init__()
|
97 |
+
|
98 |
+
self.register_modules(
|
99 |
+
unet=unet,
|
100 |
+
vae=vae,
|
101 |
+
scheduler=scheduler,
|
102 |
+
text_encoder=text_encoder,
|
103 |
+
tokenizer=tokenizer,
|
104 |
+
)
|
105 |
+
|
106 |
+
self.empty_text_embed = None
|
107 |
+
|
108 |
+
@torch.no_grad()
|
109 |
+
def __call__(
|
110 |
+
self,
|
111 |
+
input_image: Image,
|
112 |
+
denoising_steps: int = 10,
|
113 |
+
ensemble_size: int = 10,
|
114 |
+
processing_res: int = 768,
|
115 |
+
match_input_res: bool = True,
|
116 |
+
batch_size: int = 0,
|
117 |
+
color_map: str = "Spectral",
|
118 |
+
show_progress_bar: bool = True,
|
119 |
+
ensemble_kwargs: Dict = None,
|
120 |
+
) -> MarigoldDepthOutput:
|
121 |
+
"""
|
122 |
+
Function invoked when calling the pipeline.
|
123 |
+
|
124 |
+
Args:
|
125 |
+
input_image (`Image`):
|
126 |
+
Input RGB (or gray-scale) image.
|
127 |
+
processing_res (`int`, *optional*, defaults to `768`):
|
128 |
+
Maximum resolution of processing.
|
129 |
+
If set to 0: will not resize at all.
|
130 |
+
match_input_res (`bool`, *optional*, defaults to `True`):
|
131 |
+
Resize depth prediction to match input resolution.
|
132 |
+
Only valid if `limit_input_res` is not None.
|
133 |
+
denoising_steps (`int`, *optional*, defaults to `10`):
|
134 |
+
Number of diffusion denoising steps (DDIM) during inference.
|
135 |
+
ensemble_size (`int`, *optional*, defaults to `10`):
|
136 |
+
Number of predictions to be ensembled.
|
137 |
+
batch_size (`int`, *optional*, defaults to `0`):
|
138 |
+
Inference batch size, no bigger than `num_ensemble`.
|
139 |
+
If set to 0, the script will automatically decide the proper batch size.
|
140 |
+
show_progress_bar (`bool`, *optional*, defaults to `True`):
|
141 |
+
Display a progress bar of diffusion denoising.
|
142 |
+
color_map (`str`, *optional*, defaults to `"Spectral"`, pass `None` to skip colorized depth map generation):
|
143 |
+
Colormap used to colorize the depth map.
|
144 |
+
ensemble_kwargs (`dict`, *optional*, defaults to `None`):
|
145 |
+
Arguments for detailed ensembling settings.
|
146 |
+
Returns:
|
147 |
+
`MarigoldDepthOutput`: Output class for Marigold monocular depth prediction pipeline, including:
|
148 |
+
- **depth_np** (`np.ndarray`) Predicted depth map, with depth values in the range of [0, 1]
|
149 |
+
- **depth_colored** (`None` or `PIL.Image.Image`) Colorized depth map, with the shape of [3, H, W] and
|
150 |
+
values in [0, 1]. None if `color_map` is `None`
|
151 |
+
- **uncertainty** (`None` or `np.ndarray`) Uncalibrated uncertainty(MAD, median absolute deviation)
|
152 |
+
coming from ensembling. None if `ensemble_size = 1`
|
153 |
+
"""
|
154 |
+
|
155 |
+
device = self.device
|
156 |
+
input_size = input_image.size
|
157 |
+
|
158 |
+
if not match_input_res:
|
159 |
+
assert (
|
160 |
+
processing_res is not None
|
161 |
+
), "Value error: `resize_output_back` is only valid with "
|
162 |
+
assert processing_res >= 0
|
163 |
+
assert denoising_steps >= 1
|
164 |
+
assert ensemble_size >= 1
|
165 |
+
|
166 |
+
# ----------------- Image Preprocess -----------------
|
167 |
+
# Resize image
|
168 |
+
if processing_res > 0:
|
169 |
+
input_image = self.resize_max_res(
|
170 |
+
input_image, max_edge_resolution=processing_res
|
171 |
+
)
|
172 |
+
# Convert the image to RGB, to 1.remove the alpha channel 2.convert B&W to 3-channel
|
173 |
+
input_image = input_image.convert("RGB")
|
174 |
+
image = np.asarray(input_image)
|
175 |
+
|
176 |
+
# Normalize rgb values
|
177 |
+
rgb = np.transpose(image, (2, 0, 1)) # [H, W, rgb] -> [rgb, H, W]
|
178 |
+
rgb_norm = rgb / 255.0 * 2.0 - 1.0 # [0, 255] -> [-1, 1]
|
179 |
+
rgb_norm = torch.from_numpy(rgb_norm).to(self.dtype)
|
180 |
+
rgb_norm = rgb_norm.to(device)
|
181 |
+
assert rgb_norm.min() >= -1.0 and rgb_norm.max() <= 1.0
|
182 |
+
|
183 |
+
# ----------------- Predicting depth -----------------
|
184 |
+
# Batch repeated input image
|
185 |
+
duplicated_rgb = torch.stack([rgb_norm] * ensemble_size)
|
186 |
+
single_rgb_dataset = TensorDataset(duplicated_rgb)
|
187 |
+
if batch_size > 0:
|
188 |
+
_bs = batch_size
|
189 |
+
else:
|
190 |
+
_bs = self._find_batch_size(
|
191 |
+
ensemble_size=ensemble_size,
|
192 |
+
input_res=max(rgb_norm.shape[1:]),
|
193 |
+
dtype=self.dtype,
|
194 |
+
)
|
195 |
+
|
196 |
+
single_rgb_loader = DataLoader(
|
197 |
+
single_rgb_dataset, batch_size=_bs, shuffle=False
|
198 |
+
)
|
199 |
+
|
200 |
+
# Predict depth maps (batched)
|
201 |
+
depth_pred_ls = []
|
202 |
+
if show_progress_bar:
|
203 |
+
iterable = tqdm(
|
204 |
+
single_rgb_loader, desc=" " * 2 + "Inference batches", leave=False
|
205 |
+
)
|
206 |
+
else:
|
207 |
+
iterable = single_rgb_loader
|
208 |
+
for batch in iterable:
|
209 |
+
(batched_img,) = batch
|
210 |
+
depth_pred_raw = self.single_infer(
|
211 |
+
rgb_in=batched_img,
|
212 |
+
num_inference_steps=denoising_steps,
|
213 |
+
show_pbar=show_progress_bar,
|
214 |
+
)
|
215 |
+
depth_pred_ls.append(depth_pred_raw.detach().clone())
|
216 |
+
depth_preds = torch.concat(depth_pred_ls, axis=0).squeeze()
|
217 |
+
torch.cuda.empty_cache() # clear vram cache for ensembling
|
218 |
+
|
219 |
+
# ----------------- Test-time ensembling -----------------
|
220 |
+
if ensemble_size > 1:
|
221 |
+
depth_pred, pred_uncert = self.ensemble_depths(
|
222 |
+
depth_preds, **(ensemble_kwargs or {})
|
223 |
+
)
|
224 |
+
else:
|
225 |
+
depth_pred = depth_preds
|
226 |
+
pred_uncert = None
|
227 |
+
|
228 |
+
# ----------------- Post processing -----------------
|
229 |
+
# Scale prediction to [0, 1]
|
230 |
+
min_d = torch.min(depth_pred)
|
231 |
+
max_d = torch.max(depth_pred)
|
232 |
+
depth_pred = (depth_pred - min_d) / (max_d - min_d)
|
233 |
+
|
234 |
+
# Convert to numpy
|
235 |
+
depth_pred = depth_pred.cpu().numpy().astype(np.float32)
|
236 |
+
|
237 |
+
# Resize back to original resolution
|
238 |
+
if match_input_res:
|
239 |
+
pred_img = Image.fromarray(depth_pred)
|
240 |
+
pred_img = pred_img.resize(input_size)
|
241 |
+
depth_pred = np.asarray(pred_img)
|
242 |
+
|
243 |
+
# Clip output range
|
244 |
+
depth_pred = depth_pred.clip(0, 1)
|
245 |
+
|
246 |
+
# Colorize
|
247 |
+
if color_map is not None:
|
248 |
+
depth_colored = self.colorize_depth_maps(
|
249 |
+
depth_pred, 0, 1, cmap=color_map
|
250 |
+
).squeeze() # [3, H, W], value in (0, 1)
|
251 |
+
depth_colored = (depth_colored * 255).astype(np.uint8)
|
252 |
+
depth_colored_hwc = self.chw2hwc(depth_colored)
|
253 |
+
depth_colored_img = Image.fromarray(depth_colored_hwc)
|
254 |
+
else:
|
255 |
+
depth_colored_img = None
|
256 |
+
return MarigoldDepthOutput(
|
257 |
+
depth_np=depth_pred,
|
258 |
+
depth_colored=depth_colored_img,
|
259 |
+
uncertainty=pred_uncert,
|
260 |
+
)
|
261 |
+
|
262 |
+
def _encode_empty_text(self):
|
263 |
+
"""
|
264 |
+
Encode text embedding for empty prompt.
|
265 |
+
"""
|
266 |
+
prompt = ""
|
267 |
+
text_inputs = self.tokenizer(
|
268 |
+
prompt,
|
269 |
+
padding="do_not_pad",
|
270 |
+
max_length=self.tokenizer.model_max_length,
|
271 |
+
truncation=True,
|
272 |
+
return_tensors="pt",
|
273 |
+
)
|
274 |
+
text_input_ids = text_inputs.input_ids.to(self.text_encoder.device)
|
275 |
+
self.empty_text_embed = self.text_encoder(text_input_ids)[0].to(self.dtype)
|
276 |
+
|
277 |
+
@torch.no_grad()
|
278 |
+
def single_infer(
|
279 |
+
self, rgb_in: torch.Tensor, num_inference_steps: int, show_pbar: bool
|
280 |
+
) -> torch.Tensor:
|
281 |
+
"""
|
282 |
+
Perform an individual depth prediction without ensembling.
|
283 |
+
|
284 |
+
Args:
|
285 |
+
rgb_in (`torch.Tensor`):
|
286 |
+
Input RGB image.
|
287 |
+
num_inference_steps (`int`):
|
288 |
+
Number of diffusion denoisign steps (DDIM) during inference.
|
289 |
+
show_pbar (`bool`):
|
290 |
+
Display a progress bar of diffusion denoising.
|
291 |
+
Returns:
|
292 |
+
`torch.Tensor`: Predicted depth map.
|
293 |
+
"""
|
294 |
+
device = rgb_in.device
|
295 |
+
|
296 |
+
# Set timesteps
|
297 |
+
self.scheduler.set_timesteps(num_inference_steps, device=device)
|
298 |
+
timesteps = self.scheduler.timesteps # [T]
|
299 |
+
|
300 |
+
# Encode image
|
301 |
+
rgb_latent = self._encode_rgb(rgb_in)
|
302 |
+
|
303 |
+
# Initial depth map (noise)
|
304 |
+
depth_latent = torch.randn(
|
305 |
+
rgb_latent.shape, device=device, dtype=self.dtype
|
306 |
+
) # [B, 4, h, w]
|
307 |
+
|
308 |
+
# Batched empty text embedding
|
309 |
+
if self.empty_text_embed is None:
|
310 |
+
self._encode_empty_text()
|
311 |
+
batch_empty_text_embed = self.empty_text_embed.repeat(
|
312 |
+
(rgb_latent.shape[0], 1, 1)
|
313 |
+
) # [B, 2, 1024]
|
314 |
+
|
315 |
+
# Denoising loop
|
316 |
+
if show_pbar:
|
317 |
+
iterable = tqdm(
|
318 |
+
enumerate(timesteps),
|
319 |
+
total=len(timesteps),
|
320 |
+
leave=False,
|
321 |
+
desc=" " * 4 + "Diffusion denoising",
|
322 |
+
)
|
323 |
+
else:
|
324 |
+
iterable = enumerate(timesteps)
|
325 |
+
|
326 |
+
for i, t in iterable:
|
327 |
+
unet_input = torch.cat(
|
328 |
+
[rgb_latent, depth_latent], dim=1
|
329 |
+
) # this order is important
|
330 |
+
|
331 |
+
# predict the noise residual
|
332 |
+
noise_pred = self.unet(
|
333 |
+
unet_input, t, encoder_hidden_states=batch_empty_text_embed
|
334 |
+
).sample # [B, 4, h, w]
|
335 |
+
|
336 |
+
# compute the previous noisy sample x_t -> x_t-1
|
337 |
+
depth_latent = self.scheduler.step(noise_pred, t, depth_latent).prev_sample
|
338 |
+
torch.cuda.empty_cache()
|
339 |
+
depth = self._decode_depth(depth_latent)
|
340 |
+
|
341 |
+
# clip prediction
|
342 |
+
depth = torch.clip(depth, -1.0, 1.0)
|
343 |
+
# shift to [0, 1]
|
344 |
+
depth = (depth + 1.0) / 2.0
|
345 |
+
|
346 |
+
return depth
|
347 |
+
|
348 |
+
def _encode_rgb(self, rgb_in: torch.Tensor) -> torch.Tensor:
|
349 |
+
"""
|
350 |
+
Encode RGB image into latent.
|
351 |
+
|
352 |
+
Args:
|
353 |
+
rgb_in (`torch.Tensor`):
|
354 |
+
Input RGB image to be encoded.
|
355 |
+
|
356 |
+
Returns:
|
357 |
+
`torch.Tensor`: Image latent.
|
358 |
+
"""
|
359 |
+
# encode
|
360 |
+
h = self.vae.encoder(rgb_in)
|
361 |
+
moments = self.vae.quant_conv(h)
|
362 |
+
mean, logvar = torch.chunk(moments, 2, dim=1)
|
363 |
+
# scale latent
|
364 |
+
rgb_latent = mean * self.rgb_latent_scale_factor
|
365 |
+
return rgb_latent
|
366 |
+
|
367 |
+
def _decode_depth(self, depth_latent: torch.Tensor) -> torch.Tensor:
|
368 |
+
"""
|
369 |
+
Decode depth latent into depth map.
|
370 |
+
|
371 |
+
Args:
|
372 |
+
depth_latent (`torch.Tensor`):
|
373 |
+
Depth latent to be decoded.
|
374 |
+
|
375 |
+
Returns:
|
376 |
+
`torch.Tensor`: Decoded depth map.
|
377 |
+
"""
|
378 |
+
# scale latent
|
379 |
+
depth_latent = depth_latent / self.depth_latent_scale_factor
|
380 |
+
# decode
|
381 |
+
z = self.vae.post_quant_conv(depth_latent)
|
382 |
+
stacked = self.vae.decoder(z)
|
383 |
+
# mean of output channels
|
384 |
+
depth_mean = stacked.mean(dim=1, keepdim=True)
|
385 |
+
return depth_mean
|
386 |
+
|
387 |
+
@staticmethod
|
388 |
+
def resize_max_res(img: Image.Image, max_edge_resolution: int) -> Image.Image:
|
389 |
+
"""
|
390 |
+
Resize image to limit maximum edge length while keeping aspect ratio.
|
391 |
+
|
392 |
+
Args:
|
393 |
+
img (`Image.Image`):
|
394 |
+
Image to be resized.
|
395 |
+
max_edge_resolution (`int`):
|
396 |
+
Maximum edge length (pixel).
|
397 |
+
|
398 |
+
Returns:
|
399 |
+
`Image.Image`: Resized image.
|
400 |
+
"""
|
401 |
+
original_width, original_height = img.size
|
402 |
+
downscale_factor = min(
|
403 |
+
max_edge_resolution / original_width, max_edge_resolution / original_height
|
404 |
+
)
|
405 |
+
|
406 |
+
new_width = int(original_width * downscale_factor)
|
407 |
+
new_height = int(original_height * downscale_factor)
|
408 |
+
|
409 |
+
resized_img = img.resize((new_width, new_height))
|
410 |
+
return resized_img
|
411 |
+
|
412 |
+
@staticmethod
|
413 |
+
def colorize_depth_maps(
|
414 |
+
depth_map, min_depth, max_depth, cmap="Spectral", valid_mask=None
|
415 |
+
):
|
416 |
+
"""
|
417 |
+
Colorize depth maps.
|
418 |
+
"""
|
419 |
+
assert len(depth_map.shape) >= 2, "Invalid dimension"
|
420 |
+
|
421 |
+
if isinstance(depth_map, torch.Tensor):
|
422 |
+
depth = depth_map.detach().clone().squeeze().numpy()
|
423 |
+
elif isinstance(depth_map, np.ndarray):
|
424 |
+
depth = depth_map.copy().squeeze()
|
425 |
+
# reshape to [ (B,) H, W ]
|
426 |
+
if depth.ndim < 3:
|
427 |
+
depth = depth[np.newaxis, :, :]
|
428 |
+
|
429 |
+
# colorize
|
430 |
+
cm = matplotlib.colormaps[cmap]
|
431 |
+
depth = ((depth - min_depth) / (max_depth - min_depth)).clip(0, 1)
|
432 |
+
img_colored_np = cm(depth, bytes=False)[:, :, :, 0:3] # value from 0 to 1
|
433 |
+
img_colored_np = np.rollaxis(img_colored_np, 3, 1)
|
434 |
+
|
435 |
+
if valid_mask is not None:
|
436 |
+
if isinstance(depth_map, torch.Tensor):
|
437 |
+
valid_mask = valid_mask.detach().numpy()
|
438 |
+
valid_mask = valid_mask.squeeze() # [H, W] or [B, H, W]
|
439 |
+
if valid_mask.ndim < 3:
|
440 |
+
valid_mask = valid_mask[np.newaxis, np.newaxis, :, :]
|
441 |
+
else:
|
442 |
+
valid_mask = valid_mask[:, np.newaxis, :, :]
|
443 |
+
valid_mask = np.repeat(valid_mask, 3, axis=1)
|
444 |
+
img_colored_np[~valid_mask] = 0
|
445 |
+
|
446 |
+
if isinstance(depth_map, torch.Tensor):
|
447 |
+
img_colored = torch.from_numpy(img_colored_np).float()
|
448 |
+
elif isinstance(depth_map, np.ndarray):
|
449 |
+
img_colored = img_colored_np
|
450 |
+
|
451 |
+
return img_colored
|
452 |
+
|
453 |
+
@staticmethod
|
454 |
+
def chw2hwc(chw):
|
455 |
+
assert 3 == len(chw.shape)
|
456 |
+
if isinstance(chw, torch.Tensor):
|
457 |
+
hwc = torch.permute(chw, (1, 2, 0))
|
458 |
+
elif isinstance(chw, np.ndarray):
|
459 |
+
hwc = np.moveaxis(chw, 0, -1)
|
460 |
+
return hwc
|
461 |
+
|
462 |
+
@staticmethod
|
463 |
+
def _find_batch_size(ensemble_size: int, input_res: int, dtype: torch.dtype) -> int:
|
464 |
+
"""
|
465 |
+
Automatically search for suitable operating batch size.
|
466 |
+
|
467 |
+
Args:
|
468 |
+
ensemble_size (`int`):
|
469 |
+
Number of predictions to be ensembled.
|
470 |
+
input_res (`int`):
|
471 |
+
Operating resolution of the input image.
|
472 |
+
|
473 |
+
Returns:
|
474 |
+
`int`: Operating batch size.
|
475 |
+
"""
|
476 |
+
# Search table for suggested max. inference batch size
|
477 |
+
bs_search_table = [
|
478 |
+
# tested on A100-PCIE-80GB
|
479 |
+
{"res": 768, "total_vram": 79, "bs": 35, "dtype": torch.float32},
|
480 |
+
{"res": 1024, "total_vram": 79, "bs": 20, "dtype": torch.float32},
|
481 |
+
# tested on A100-PCIE-40GB
|
482 |
+
{"res": 768, "total_vram": 39, "bs": 15, "dtype": torch.float32},
|
483 |
+
{"res": 1024, "total_vram": 39, "bs": 8, "dtype": torch.float32},
|
484 |
+
{"res": 768, "total_vram": 39, "bs": 30, "dtype": torch.float16},
|
485 |
+
{"res": 1024, "total_vram": 39, "bs": 15, "dtype": torch.float16},
|
486 |
+
# tested on RTX3090, RTX4090
|
487 |
+
{"res": 512, "total_vram": 23, "bs": 20, "dtype": torch.float32},
|
488 |
+
{"res": 768, "total_vram": 23, "bs": 7, "dtype": torch.float32},
|
489 |
+
{"res": 1024, "total_vram": 23, "bs": 3, "dtype": torch.float32},
|
490 |
+
{"res": 512, "total_vram": 23, "bs": 40, "dtype": torch.float16},
|
491 |
+
{"res": 768, "total_vram": 23, "bs": 18, "dtype": torch.float16},
|
492 |
+
{"res": 1024, "total_vram": 23, "bs": 10, "dtype": torch.float16},
|
493 |
+
# tested on GTX1080Ti
|
494 |
+
{"res": 512, "total_vram": 10, "bs": 5, "dtype": torch.float32},
|
495 |
+
{"res": 768, "total_vram": 10, "bs": 2, "dtype": torch.float32},
|
496 |
+
{"res": 512, "total_vram": 10, "bs": 10, "dtype": torch.float16},
|
497 |
+
{"res": 768, "total_vram": 10, "bs": 5, "dtype": torch.float16},
|
498 |
+
{"res": 1024, "total_vram": 10, "bs": 3, "dtype": torch.float16},
|
499 |
+
]
|
500 |
+
|
501 |
+
if not torch.cuda.is_available():
|
502 |
+
return 1
|
503 |
+
|
504 |
+
total_vram = torch.cuda.mem_get_info()[1] / 1024.0**3
|
505 |
+
filtered_bs_search_table = [s for s in bs_search_table if s["dtype"] == dtype]
|
506 |
+
for settings in sorted(
|
507 |
+
filtered_bs_search_table,
|
508 |
+
key=lambda k: (k["res"], -k["total_vram"]),
|
509 |
+
):
|
510 |
+
if input_res <= settings["res"] and total_vram >= settings["total_vram"]:
|
511 |
+
bs = settings["bs"]
|
512 |
+
if bs > ensemble_size:
|
513 |
+
bs = ensemble_size
|
514 |
+
elif bs > math.ceil(ensemble_size / 2) and bs < ensemble_size:
|
515 |
+
bs = math.ceil(ensemble_size / 2)
|
516 |
+
return bs
|
517 |
+
|
518 |
+
return 1
|
519 |
+
|
520 |
+
@staticmethod
|
521 |
+
def ensemble_depths(
|
522 |
+
input_images: torch.Tensor,
|
523 |
+
regularizer_strength: float = 0.02,
|
524 |
+
max_iter: int = 2,
|
525 |
+
tol: float = 1e-3,
|
526 |
+
reduction: str = "median",
|
527 |
+
max_res: int = None,
|
528 |
+
):
|
529 |
+
"""
|
530 |
+
To ensemble multiple affine-invariant depth images (up to scale and shift),
|
531 |
+
by aligning estimating the scale and shift
|
532 |
+
"""
|
533 |
+
|
534 |
+
def inter_distances(tensors: torch.Tensor):
|
535 |
+
"""
|
536 |
+
To calculate the distance between each two depth maps.
|
537 |
+
"""
|
538 |
+
distances = []
|
539 |
+
for i, j in torch.combinations(torch.arange(tensors.shape[0])):
|
540 |
+
arr1 = tensors[i : i + 1]
|
541 |
+
arr2 = tensors[j : j + 1]
|
542 |
+
distances.append(arr1 - arr2)
|
543 |
+
dist = torch.concatenate(distances, dim=0)
|
544 |
+
return dist
|
545 |
+
|
546 |
+
device = input_images.device
|
547 |
+
dtype = input_images.dtype
|
548 |
+
np_dtype = np.float32
|
549 |
+
|
550 |
+
original_input = input_images.clone()
|
551 |
+
n_img = input_images.shape[0]
|
552 |
+
ori_shape = input_images.shape
|
553 |
+
|
554 |
+
if max_res is not None:
|
555 |
+
scale_factor = torch.min(max_res / torch.tensor(ori_shape[-2:]))
|
556 |
+
if scale_factor < 1:
|
557 |
+
downscaler = torch.nn.Upsample(
|
558 |
+
scale_factor=scale_factor, mode="nearest"
|
559 |
+
)
|
560 |
+
input_images = downscaler(torch.from_numpy(input_images)).numpy()
|
561 |
+
|
562 |
+
# init guess
|
563 |
+
_min = np.min(input_images.reshape((n_img, -1)).cpu().numpy(), axis=1)
|
564 |
+
_max = np.max(input_images.reshape((n_img, -1)).cpu().numpy(), axis=1)
|
565 |
+
s_init = 1.0 / (_max - _min).reshape((-1, 1, 1))
|
566 |
+
t_init = (-1 * s_init.flatten() * _min.flatten()).reshape((-1, 1, 1))
|
567 |
+
x = np.concatenate([s_init, t_init]).reshape(-1).astype(np_dtype)
|
568 |
+
|
569 |
+
input_images = input_images.to(device)
|
570 |
+
|
571 |
+
# objective function
|
572 |
+
def closure(x):
|
573 |
+
l = len(x)
|
574 |
+
s = x[: int(l / 2)]
|
575 |
+
t = x[int(l / 2) :]
|
576 |
+
s = torch.from_numpy(s).to(dtype=dtype).to(device)
|
577 |
+
t = torch.from_numpy(t).to(dtype=dtype).to(device)
|
578 |
+
|
579 |
+
transformed_arrays = input_images * s.view((-1, 1, 1)) + t.view((-1, 1, 1))
|
580 |
+
dists = inter_distances(transformed_arrays)
|
581 |
+
sqrt_dist = torch.sqrt(torch.mean(dists**2))
|
582 |
+
|
583 |
+
if "mean" == reduction:
|
584 |
+
pred = torch.mean(transformed_arrays, dim=0)
|
585 |
+
elif "median" == reduction:
|
586 |
+
pred = torch.median(transformed_arrays, dim=0).values
|
587 |
+
else:
|
588 |
+
raise ValueError
|
589 |
+
|
590 |
+
near_err = torch.sqrt((0 - torch.min(pred)) ** 2)
|
591 |
+
far_err = torch.sqrt((1 - torch.max(pred)) ** 2)
|
592 |
+
|
593 |
+
err = sqrt_dist + (near_err + far_err) * regularizer_strength
|
594 |
+
err = err.detach().cpu().numpy().astype(np_dtype)
|
595 |
+
return err
|
596 |
+
|
597 |
+
res = minimize(
|
598 |
+
closure,
|
599 |
+
x,
|
600 |
+
method="BFGS",
|
601 |
+
tol=tol,
|
602 |
+
options={"maxiter": max_iter, "disp": False},
|
603 |
+
)
|
604 |
+
x = res.x
|
605 |
+
l = len(x)
|
606 |
+
s = x[: int(l / 2)]
|
607 |
+
t = x[int(l / 2) :]
|
608 |
+
|
609 |
+
# Prediction
|
610 |
+
s = torch.from_numpy(s).to(dtype=dtype).to(device)
|
611 |
+
t = torch.from_numpy(t).to(dtype=dtype).to(device)
|
612 |
+
transformed_arrays = original_input * s.view(-1, 1, 1) + t.view(-1, 1, 1)
|
613 |
+
if "mean" == reduction:
|
614 |
+
aligned_images = torch.mean(transformed_arrays, dim=0)
|
615 |
+
std = torch.std(transformed_arrays, dim=0)
|
616 |
+
uncertainty = std
|
617 |
+
elif "median" == reduction:
|
618 |
+
aligned_images = torch.median(transformed_arrays, dim=0).values
|
619 |
+
# MAD (median absolute deviation) as uncertainty indicator
|
620 |
+
abs_dev = torch.abs(transformed_arrays - aligned_images)
|
621 |
+
mad = torch.median(abs_dev, dim=0).values
|
622 |
+
uncertainty = mad
|
623 |
+
else:
|
624 |
+
raise ValueError(f"Unknown reduction method: {reduction}")
|
625 |
+
|
626 |
+
# Scale and shift to [0, 1]
|
627 |
+
_min = torch.min(aligned_images)
|
628 |
+
_max = torch.max(aligned_images)
|
629 |
+
aligned_images = (aligned_images - _min) / (_max - _min)
|
630 |
+
uncertainty /= _max - _min
|
631 |
+
|
632 |
+
return aligned_images, uncertainty
|
requirements.txt
CHANGED
@@ -1,13 +1,13 @@
|
|
1 |
-
gradio==4.
|
2 |
gradio-imageslider==0.0.16
|
3 |
-
GitPython==3.1.40
|
4 |
pygltflib==1.16.1
|
5 |
trimesh==4.0.5
|
6 |
|
|
|
7 |
accelerate>=0.22.0
|
8 |
-
diffusers
|
9 |
matplotlib==3.8.2
|
10 |
scipy==1.11.4
|
11 |
torch==2.0.1
|
12 |
transformers>=4.32.1
|
13 |
-
xformers
|
|
|
1 |
+
gradio==4.21.0
|
2 |
gradio-imageslider==0.0.16
|
|
|
3 |
pygltflib==1.16.1
|
4 |
trimesh==4.0.5
|
5 |
|
6 |
+
spaces>=0.25.0
|
7 |
accelerate>=0.22.0
|
8 |
+
diffusers==0.27.2
|
9 |
matplotlib==3.8.2
|
10 |
scipy==1.11.4
|
11 |
torch==2.0.1
|
12 |
transformers>=4.32.1
|
13 |
+
xformers>=0.0.21
|