Jiaye Zou
commited on
Commit
•
50318d8
1
Parent(s):
f474bfd
update: gradio app with docker
Browse files- Dockerfile +30 -0
- README.md +2 -3
- app.py +62 -12
- config.yaml +36 -0
- get_weights.sh +9 -0
- mapper/utils/viz_2d.py +43 -13
- requirements.txt +23 -0
Dockerfile
ADDED
@@ -0,0 +1,30 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
FROM pytorch/pytorch:2.3.1-cuda11.8-cudnn8-runtime
|
2 |
+
|
3 |
+
# Set working directory
|
4 |
+
WORKDIR /mapper
|
5 |
+
|
6 |
+
# Install dependencies
|
7 |
+
RUN apt-get update && apt-get install -y \
|
8 |
+
git \
|
9 |
+
wget \
|
10 |
+
unzip \
|
11 |
+
vim \
|
12 |
+
ffmpeg \
|
13 |
+
libsm6 \
|
14 |
+
libxext6
|
15 |
+
|
16 |
+
RUN pip install --no-cache-dir gradio[oauth]==4.36.1 "uvicorn>=0.14.0" spaces
|
17 |
+
|
18 |
+
COPY . /mapper
|
19 |
+
|
20 |
+
# Install Python dependencies
|
21 |
+
RUN pip install --no-cache-dir -r requirements.txt
|
22 |
+
|
23 |
+
# Get Weights
|
24 |
+
RUN bash get_weights.sh
|
25 |
+
|
26 |
+
# Clear APT and pip cache
|
27 |
+
RUN apt-get clean && rm -rf /var/lib/apt/lists/* && rm -rf /tmp/pip-reqs
|
28 |
+
|
29 |
+
# Start the app
|
30 |
+
CMD ["python", "app.py"]
|
README.md
CHANGED
@@ -3,10 +3,9 @@ title: "Map It Anywhere (MIA): Empowering Bird’s Eye View Mapping using Large-
|
|
3 |
emoji: 🌍
|
4 |
colorFrom: green
|
5 |
colorTo: blue
|
6 |
-
sdk:
|
7 |
-
sdk_version: "4.36.1"
|
8 |
-
app_file: app.py
|
9 |
pinned: true
|
|
|
10 |
---
|
11 |
<p align="center">
|
12 |
<h1 align="center">Map It Anywhere (MIA): Empowering Bird’s Eye View Mapping using Large-scale Public Data</h1>
|
|
|
3 |
emoji: 🌍
|
4 |
colorFrom: green
|
5 |
colorTo: blue
|
6 |
+
sdk: docker
|
|
|
|
|
7 |
pinned: true
|
8 |
+
app_port: 7860
|
9 |
---
|
10 |
<p align="center">
|
11 |
<h1 align="center">Map It Anywhere (MIA): Empowering Bird’s Eye View Mapping using Large-scale Public Data</h1>
|
app.py
CHANGED
@@ -3,9 +3,14 @@ from matplotlib import pyplot as plt
|
|
3 |
from mapper.utils.io import read_image
|
4 |
from mapper.utils.exif import EXIF
|
5 |
from mapper.utils.wrappers import Camera
|
|
|
|
|
|
|
6 |
from perspective2d import PerspectiveFields
|
|
|
7 |
import numpy as np
|
8 |
from typing import Optional, Tuple
|
|
|
9 |
|
10 |
description = """
|
11 |
<h1 align="center">
|
@@ -24,6 +29,10 @@ Mapper generates birds-eye-view maps from first person view monocular images. Tr
|
|
24 |
</p>
|
25 |
"""
|
26 |
|
|
|
|
|
|
|
|
|
27 |
class ImageCalibrator(PerspectiveFields):
|
28 |
def __init__(self, version: str = "Paramnet-360Cities-edina-centered"):
|
29 |
super().__init__(version)
|
@@ -40,7 +49,6 @@ class ImageCalibrator(PerspectiveFields):
|
|
40 |
_, focal_ratio = exif.extract_focal()
|
41 |
if focal_ratio != 0:
|
42 |
focal_length = focal_ratio * max(h, w)
|
43 |
-
|
44 |
calib = self.inference(img_bgr=image_rgb[..., ::-1])
|
45 |
roll_pitch = (calib["pred_roll"].item(), calib["pred_pitch"].item())
|
46 |
if focal_length is None:
|
@@ -57,26 +65,67 @@ class ImageCalibrator(PerspectiveFields):
|
|
57 |
)
|
58 |
return roll_pitch, camera
|
59 |
|
60 |
-
def
|
61 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
62 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
63 |
image_path = input_img.name
|
64 |
|
65 |
image = read_image(image_path)
|
66 |
-
image = image.to("cuda")
|
67 |
with open(image_path, "rb") as fid:
|
68 |
exif = EXIF(fid, lambda: image.shape[:2])
|
69 |
|
70 |
gravity, camera = calibrator.run(image, exif=exif)
|
71 |
|
72 |
-
|
73 |
-
|
74 |
-
|
75 |
-
|
76 |
-
|
77 |
fig1 = plt.gcf()
|
78 |
|
79 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
80 |
|
81 |
demo = gr.Interface(
|
82 |
fn=run,
|
@@ -84,7 +133,8 @@ demo = gr.Interface(
|
|
84 |
gr.File(file_types=["image"], label="Input Image")
|
85 |
],
|
86 |
outputs=[
|
87 |
-
gr.Plot(label="Inputs", format="png")
|
|
|
88 |
],
|
89 |
description=description,)
|
90 |
-
demo.launch(share=
|
|
|
3 |
from mapper.utils.io import read_image
|
4 |
from mapper.utils.exif import EXIF
|
5 |
from mapper.utils.wrappers import Camera
|
6 |
+
from mapper.data.image import rectify_image, pad_image, resize_image
|
7 |
+
from mapper.utils.viz_2d import one_hot_argmax_to_rgb, plot_images
|
8 |
+
from mapper.module import GenericModule
|
9 |
from perspective2d import PerspectiveFields
|
10 |
+
import torch
|
11 |
import numpy as np
|
12 |
from typing import Optional, Tuple
|
13 |
+
from omegaconf import OmegaConf
|
14 |
|
15 |
description = """
|
16 |
<h1 align="center">
|
|
|
29 |
</p>
|
30 |
"""
|
31 |
|
32 |
+
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
33 |
+
|
34 |
+
cfg = OmegaConf.load("config.yaml")
|
35 |
+
|
36 |
class ImageCalibrator(PerspectiveFields):
|
37 |
def __init__(self, version: str = "Paramnet-360Cities-edina-centered"):
|
38 |
super().__init__(version)
|
|
|
49 |
_, focal_ratio = exif.extract_focal()
|
50 |
if focal_ratio != 0:
|
51 |
focal_length = focal_ratio * max(h, w)
|
|
|
52 |
calib = self.inference(img_bgr=image_rgb[..., ::-1])
|
53 |
roll_pitch = (calib["pred_roll"].item(), calib["pred_pitch"].item())
|
54 |
if focal_length is None:
|
|
|
65 |
)
|
66 |
return roll_pitch, camera
|
67 |
|
68 |
+
def preprocess_pipeline(image, roll_pitch, camera):
|
69 |
+
image = torch.from_numpy(image).float() / 255
|
70 |
+
image = image.permute(2, 0, 1).to(device)
|
71 |
+
camera = camera.to(device)
|
72 |
+
|
73 |
+
image, valid = rectify_image(image, camera.float(), -roll_pitch[0], -roll_pitch[1])
|
74 |
+
|
75 |
+
roll_pitch *= 0
|
76 |
+
|
77 |
+
image, _, camera, valid = resize_image(
|
78 |
+
image=image,
|
79 |
+
size=512,
|
80 |
+
camera=camera,
|
81 |
+
fn=max,
|
82 |
+
valid=valid
|
83 |
+
)
|
84 |
+
|
85 |
+
image, valid, camera = pad_image(
|
86 |
+
image, 512, camera, valid
|
87 |
+
)
|
88 |
|
89 |
+
camera = torch.stack([camera])
|
90 |
+
|
91 |
+
return {
|
92 |
+
"image": image.unsqueeze(0).to(device),
|
93 |
+
"valid": valid.unsqueeze(0).to(device),
|
94 |
+
"camera": camera.float().to(device),
|
95 |
+
}
|
96 |
+
|
97 |
+
|
98 |
+
calibrator = ImageCalibrator().to(device)
|
99 |
+
model = GenericModule(cfg)
|
100 |
+
model = model.load_from_checkpoint("trained_weights/mapper-excl-ood.ckpt", strict=False, cfg=cfg)
|
101 |
+
model = model.to(device)
|
102 |
+
model = model.eval()
|
103 |
+
|
104 |
+
def run(input_img):
|
105 |
image_path = input_img.name
|
106 |
|
107 |
image = read_image(image_path)
|
|
|
108 |
with open(image_path, "rb") as fid:
|
109 |
exif = EXIF(fid, lambda: image.shape[:2])
|
110 |
|
111 |
gravity, camera = calibrator.run(image, exif=exif)
|
112 |
|
113 |
+
data = preprocess_pipeline(image, gravity, camera)
|
114 |
+
|
115 |
+
res = model(data)
|
116 |
+
|
117 |
+
plot_images([image], pad=0., adaptive=True)
|
118 |
fig1 = plt.gcf()
|
119 |
|
120 |
+
prediction = res['output']
|
121 |
+
rgb_prediction = one_hot_argmax_to_rgb(prediction, 6).squeeze(0).permute(1, 2, 0).cpu().long().numpy()
|
122 |
+
valid = res['valid_bev'].squeeze(0)[..., :-1]
|
123 |
+
rgb_prediction[~valid.cpu().numpy()] = 255
|
124 |
+
|
125 |
+
plot_images([rgb_prediction], pad=0., adaptive=True)
|
126 |
+
fig2 = plt.gcf()
|
127 |
+
|
128 |
+
return fig1, fig2
|
129 |
|
130 |
demo = gr.Interface(
|
131 |
fn=run,
|
|
|
133 |
gr.File(file_types=["image"], label="Input Image")
|
134 |
],
|
135 |
outputs=[
|
136 |
+
gr.Plot(label="Inputs", format="png"),
|
137 |
+
gr.Plot(label="Outputs", format="png"),
|
138 |
],
|
139 |
description=description,)
|
140 |
+
demo.launch(share=False, server_name="0.0.0.0")
|
config.yaml
ADDED
@@ -0,0 +1,36 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
model:
|
2 |
+
image_encoder:
|
3 |
+
backbone:
|
4 |
+
pretrained: true
|
5 |
+
frozen: true
|
6 |
+
output_dim: 128
|
7 |
+
name: feature_extractor_DPT
|
8 |
+
segmentation_head:
|
9 |
+
dropout_rate: 0.2
|
10 |
+
name: map_perception_net
|
11 |
+
num_classes: 6
|
12 |
+
latent_dim: 128
|
13 |
+
z_max: 50
|
14 |
+
x_max: 25
|
15 |
+
pixel_per_meter: 2
|
16 |
+
num_scale_bins: 32
|
17 |
+
loss:
|
18 |
+
num_classes: 6
|
19 |
+
xent_weight: 1.0
|
20 |
+
dice_weight: 1.0
|
21 |
+
focal_loss: false
|
22 |
+
focal_loss_gamma: 2.0
|
23 |
+
requires_frustrum: true
|
24 |
+
requires_flood_mask: false
|
25 |
+
class_weights:
|
26 |
+
- 1.00351229
|
27 |
+
- 4.34782609
|
28 |
+
- 1.00110121
|
29 |
+
- 1.03124678
|
30 |
+
- 6.69792364
|
31 |
+
- 7.55857899
|
32 |
+
label_smoothing: 0.1
|
33 |
+
scale_range:
|
34 |
+
- 0
|
35 |
+
- 9
|
36 |
+
z_min: null
|
get_weights.sh
ADDED
@@ -0,0 +1,9 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
#!/bin/bash
|
2 |
+
|
3 |
+
# URL of the file to download
|
4 |
+
ood_weights="https://huggingface.co/mapitanywhere/mapper/resolve/main/weights/mapper-excl-ood/model.ckpt"
|
5 |
+
|
6 |
+
mkdir -p trained_weights
|
7 |
+
|
8 |
+
# Download the file using curl
|
9 |
+
wget $ood_weights -O trained_weights/mapper-excl-ood.ckpt
|
mapper/utils/viz_2d.py
CHANGED
@@ -6,6 +6,7 @@
|
|
6 |
|
7 |
import numpy as np
|
8 |
import torch
|
|
|
9 |
|
10 |
|
11 |
def features_to_RGB(*Fs, masks=None, skip=1):
|
@@ -69,22 +70,18 @@ def one_hot_argmax_to_rgb(y, num_class):
|
|
69 |
|
70 |
'''
|
71 |
|
72 |
-
|
73 |
class_colors = {
|
74 |
-
'road': (
|
75 |
-
'crossing': (
|
76 |
-
'explicit_pedestrian': (
|
77 |
# 'explicit_void': (128, 128, 128), # 3: White
|
78 |
-
'
|
79 |
-
'
|
80 |
-
'
|
81 |
-
'
|
82 |
-
'parking': (170, 170, 170), # 8: Dark Grey
|
83 |
-
'train': (85, 85, 85) , # 9: Light Grey
|
84 |
-
'predicted_void': (256, 256, 256)
|
85 |
}
|
86 |
class_colors = class_colors.values()
|
87 |
-
class_colors = [torch.tensor(x) for x in class_colors]
|
88 |
|
89 |
argmaxed = torch.argmax((y > 0.5).float(), dim=1) # Take argmax
|
90 |
argmaxed[torch.all(y <= 0.5, dim=1)] = num_class
|
@@ -97,10 +94,43 @@ def one_hot_argmax_to_rgb(y, num_class):
|
|
97 |
argmaxed.shape[1],
|
98 |
argmaxed.shape[2],
|
99 |
)
|
100 |
-
) *
|
101 |
for i in range(num_class + 1):
|
102 |
seg_rgb[:, 0, :, :][argmaxed == i] = class_colors[i][0]
|
103 |
seg_rgb[:, 1, :, :][argmaxed == i] = class_colors[i][1]
|
104 |
seg_rgb[:, 2, :, :][argmaxed == i] = class_colors[i][2]
|
105 |
|
106 |
return seg_rgb
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
6 |
|
7 |
import numpy as np
|
8 |
import torch
|
9 |
+
import matplotlib.pyplot as plt
|
10 |
|
11 |
|
12 |
def features_to_RGB(*Fs, masks=None, skip=1):
|
|
|
70 |
|
71 |
'''
|
72 |
|
|
|
73 |
class_colors = {
|
74 |
+
'road': (68, 68, 68), # 0: Black
|
75 |
+
'crossing': (244, 162, 97), # 1; Red
|
76 |
+
'explicit_pedestrian': (233, 196, 106), # 2: Yellow
|
77 |
# 'explicit_void': (128, 128, 128), # 3: White
|
78 |
+
'building': (231, 111, 81), # 5: Magenta
|
79 |
+
'terrain': (42, 157, 143), # 7: Cyan
|
80 |
+
'parking': (204, 204, 204), # 8: Dark Grey
|
81 |
+
'predicted_void': (255, 255, 255)
|
|
|
|
|
|
|
82 |
}
|
83 |
class_colors = class_colors.values()
|
84 |
+
class_colors = [torch.tensor(x).float() for x in class_colors]
|
85 |
|
86 |
argmaxed = torch.argmax((y > 0.5).float(), dim=1) # Take argmax
|
87 |
argmaxed[torch.all(y <= 0.5, dim=1)] = num_class
|
|
|
94 |
argmaxed.shape[1],
|
95 |
argmaxed.shape[2],
|
96 |
)
|
97 |
+
) * 255
|
98 |
for i in range(num_class + 1):
|
99 |
seg_rgb[:, 0, :, :][argmaxed == i] = class_colors[i][0]
|
100 |
seg_rgb[:, 1, :, :][argmaxed == i] = class_colors[i][1]
|
101 |
seg_rgb[:, 2, :, :][argmaxed == i] = class_colors[i][2]
|
102 |
|
103 |
return seg_rgb
|
104 |
+
|
105 |
+
def plot_images(imgs, titles=None, cmaps="gray", dpi=100, pad=0.5, adaptive=True):
|
106 |
+
"""Plot a set of images horizontally.
|
107 |
+
Args:
|
108 |
+
imgs: a list of NumPy or PyTorch images, RGB (H, W, 3) or mono (H, W).
|
109 |
+
titles: a list of strings, as titles for each image.
|
110 |
+
cmaps: colormaps for monochrome images.
|
111 |
+
adaptive: whether the figure size should fit the image aspect ratios.
|
112 |
+
"""
|
113 |
+
n = len(imgs)
|
114 |
+
if not isinstance(cmaps, (list, tuple)):
|
115 |
+
cmaps = [cmaps] * n
|
116 |
+
|
117 |
+
if adaptive:
|
118 |
+
ratios = [i.shape[1] / i.shape[0] for i in imgs] # W / H
|
119 |
+
else:
|
120 |
+
ratios = [4 / 3] * n
|
121 |
+
figsize = [sum(ratios) * 4.5, 4.5]
|
122 |
+
fig, ax = plt.subplots(
|
123 |
+
1, n, figsize=figsize, dpi=dpi, gridspec_kw={"width_ratios": ratios}
|
124 |
+
)
|
125 |
+
if n == 1:
|
126 |
+
ax = [ax]
|
127 |
+
for i in range(n):
|
128 |
+
ax[i].imshow(imgs[i], cmap=plt.get_cmap(cmaps[i]))
|
129 |
+
ax[i].get_yaxis().set_ticks([])
|
130 |
+
ax[i].get_xaxis().set_ticks([])
|
131 |
+
ax[i].set_axis_off()
|
132 |
+
for spine in ax[i].spines.values(): # remove frame
|
133 |
+
spine.set_visible(False)
|
134 |
+
if titles:
|
135 |
+
ax[i].set_title(titles[i])
|
136 |
+
fig.tight_layout(pad=pad)
|
requirements.txt
ADDED
@@ -0,0 +1,23 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
torch
|
2 |
+
torchvision
|
3 |
+
numpy
|
4 |
+
opencv-python
|
5 |
+
Pillow
|
6 |
+
tqdm>=4.36.0
|
7 |
+
matplotlib
|
8 |
+
plotly
|
9 |
+
scipy
|
10 |
+
omegaconf
|
11 |
+
pytorch-lightning
|
12 |
+
torchmetrics
|
13 |
+
lxml
|
14 |
+
rtree
|
15 |
+
scikit-learn
|
16 |
+
geopy
|
17 |
+
exifread
|
18 |
+
hydra-core
|
19 |
+
umsgpack
|
20 |
+
nuscenes-devkit
|
21 |
+
perspective2d @ git+https://github.com/jinlinyi/PerspectiveFields.git
|
22 |
+
urllib3>=2
|
23 |
+
wandb
|