Spaces:
Running
Running
zhang-ziang
commited on
Commit
•
00fe360
1
Parent(s):
864becb
new model
Browse files- app.py +18 -63
- inference.py +49 -0
app.py
CHANGED
@@ -1,17 +1,14 @@
|
|
1 |
import gradio as gr
|
2 |
from paths import *
|
3 |
-
|
4 |
from vision_tower import DINOv2_MLP
|
5 |
from transformers import AutoImageProcessor
|
6 |
import torch
|
7 |
-
import
|
8 |
-
from PIL import Image
|
9 |
-
|
10 |
-
import torch.nn.functional as F
|
11 |
from utils import *
|
12 |
|
13 |
from huggingface_hub import hf_hub_download
|
14 |
-
ckpt_path = hf_hub_download(repo_id="Viglong/
|
15 |
print(ckpt_path)
|
16 |
|
17 |
save_path = './'
|
@@ -19,79 +16,37 @@ device = 'cpu'
|
|
19 |
dino = DINOv2_MLP(
|
20 |
dino_mode = 'large',
|
21 |
in_dim = 1024,
|
22 |
-
out_dim = 360+180+
|
23 |
evaluate = True,
|
24 |
mask_dino = False,
|
25 |
frozen_back = False
|
26 |
-
)
|
27 |
|
28 |
dino.eval()
|
29 |
print('model create')
|
30 |
dino.load_state_dict(torch.load(ckpt_path, map_location='cpu'))
|
|
|
31 |
print('weight loaded')
|
32 |
val_preprocess = AutoImageProcessor.from_pretrained(DINO_LARGE, cache_dir='./')
|
33 |
|
34 |
-
|
35 |
-
def get_3angle(image):
|
36 |
-
|
37 |
-
# image = Image.open(image_path).convert('RGB')
|
38 |
-
image_inputs = val_preprocess(images = image)
|
39 |
-
image_inputs['pixel_values'] = torch.from_numpy(np.array(image_inputs['pixel_values'])).to(device)
|
40 |
-
with torch.no_grad():
|
41 |
-
dino_pred = dino(image_inputs)
|
42 |
-
|
43 |
-
gaus_ax_pred = torch.argmax(dino_pred[:, 0:360], dim=-1)
|
44 |
-
gaus_pl_pred = torch.argmax(dino_pred[:, 360:360+180], dim=-1)
|
45 |
-
gaus_ro_pred = torch.argmax(dino_pred[:, 360+180:360+180+60], dim=-1)
|
46 |
-
confidence = F.softmax(dino_pred[:, -2:], dim=-1)[0][0]
|
47 |
-
angles = torch.zeros(4)
|
48 |
-
angles[0] = gaus_ax_pred
|
49 |
-
angles[1] = gaus_pl_pred - 90
|
50 |
-
angles[2] = gaus_ro_pred - 30
|
51 |
-
angles[3] = confidence
|
52 |
-
return angles
|
53 |
-
|
54 |
-
def get_3angle_infer_aug(origin_img, rm_bkg_img):
|
55 |
-
|
56 |
-
# image = Image.open(image_path).convert('RGB')
|
57 |
-
image = get_crop_images(origin_img, num=3) + get_crop_images(rm_bkg_img, num=3)
|
58 |
-
image_inputs = val_preprocess(images = image)
|
59 |
-
image_inputs['pixel_values'] = torch.from_numpy(np.array(image_inputs['pixel_values'])).to(device)
|
60 |
-
with torch.no_grad():
|
61 |
-
dino_pred = dino(image_inputs)
|
62 |
-
|
63 |
-
gaus_ax_pred = torch.argmax(dino_pred[:, 0:360], dim=-1).to(torch.float32)
|
64 |
-
gaus_pl_pred = torch.argmax(dino_pred[:, 360:360+180], dim=-1).to(torch.float32)
|
65 |
-
gaus_ro_pred = torch.argmax(dino_pred[:, 360+180:360+180+60], dim=-1).to(torch.float32)
|
66 |
-
|
67 |
-
gaus_ax_pred = remove_outliers_and_average_circular(gaus_ax_pred)
|
68 |
-
gaus_pl_pred = remove_outliers_and_average(gaus_pl_pred)
|
69 |
-
gaus_ro_pred = remove_outliers_and_average(gaus_ro_pred)
|
70 |
-
|
71 |
-
confidence = torch.mean(F.softmax(dino_pred[:, -2:], dim=-1), dim=0)[0]
|
72 |
-
angles = torch.zeros(4)
|
73 |
-
angles[0] = gaus_ax_pred
|
74 |
-
angles[1] = gaus_pl_pred - 90
|
75 |
-
angles[2] = gaus_ro_pred - 30
|
76 |
-
angles[3] = confidence
|
77 |
-
return angles
|
78 |
-
|
79 |
def infer_func(img, do_rm_bkg, do_infer_aug):
|
80 |
origin_img = Image.fromarray(img)
|
81 |
if do_infer_aug:
|
82 |
rm_bkg_img = background_preprocess(origin_img, True)
|
83 |
-
angles = get_3angle_infer_aug(origin_img, rm_bkg_img)
|
84 |
else:
|
85 |
rm_bkg_img = background_preprocess(origin_img, do_rm_bkg)
|
86 |
-
angles = get_3angle(rm_bkg_img)
|
87 |
|
88 |
phi = np.radians(angles[0])
|
89 |
theta = np.radians(angles[1])
|
90 |
gamma = angles[2]
|
91 |
-
|
92 |
-
|
93 |
-
|
94 |
-
|
|
|
|
|
95 |
|
96 |
# axis_model = "axis.obj"
|
97 |
return [res_img, round(float(angles[0]), 2), round(float(angles[1]), 2), round(float(angles[2]), 2), round(float(angles[3]), 2)]
|
@@ -107,10 +62,10 @@ server = gr.Interface(
|
|
107 |
outputs=[
|
108 |
gr.Image(height=512, width=512, label="result image"),
|
109 |
# gr.Model3D(clear_color=[0.0, 0.0, 0.0, 0.0], label="3D Model"),
|
110 |
-
gr.Textbox(lines=1, label='Azimuth(0~360°)'),
|
111 |
-
gr.Textbox(lines=1, label='Polar(-90~90°)'),
|
112 |
-
gr.Textbox(lines=1, label='Rotation(-90~90°)'),
|
113 |
-
gr.Textbox(lines=1, label='Confidence(0~1)')
|
114 |
]
|
115 |
)
|
116 |
|
|
|
1 |
import gradio as gr
|
2 |
from paths import *
|
3 |
+
|
4 |
from vision_tower import DINOv2_MLP
|
5 |
from transformers import AutoImageProcessor
|
6 |
import torch
|
7 |
+
from inference import *
|
|
|
|
|
|
|
8 |
from utils import *
|
9 |
|
10 |
from huggingface_hub import hf_hub_download
|
11 |
+
ckpt_path = hf_hub_download(repo_id="Viglong/Orient-Anything", filename="croplargeEX2/dino_weight.pt", repo_type="model", cache_dir='./', resume_download=True)
|
12 |
print(ckpt_path)
|
13 |
|
14 |
save_path = './'
|
|
|
16 |
dino = DINOv2_MLP(
|
17 |
dino_mode = 'large',
|
18 |
in_dim = 1024,
|
19 |
+
out_dim = 360+180+180+2,
|
20 |
evaluate = True,
|
21 |
mask_dino = False,
|
22 |
frozen_back = False
|
23 |
+
)
|
24 |
|
25 |
dino.eval()
|
26 |
print('model create')
|
27 |
dino.load_state_dict(torch.load(ckpt_path, map_location='cpu'))
|
28 |
+
dino = dino.to(device)
|
29 |
print('weight loaded')
|
30 |
val_preprocess = AutoImageProcessor.from_pretrained(DINO_LARGE, cache_dir='./')
|
31 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
32 |
def infer_func(img, do_rm_bkg, do_infer_aug):
|
33 |
origin_img = Image.fromarray(img)
|
34 |
if do_infer_aug:
|
35 |
rm_bkg_img = background_preprocess(origin_img, True)
|
36 |
+
angles = get_3angle_infer_aug(origin_img, rm_bkg_img, dino, val_preprocess, device)
|
37 |
else:
|
38 |
rm_bkg_img = background_preprocess(origin_img, do_rm_bkg)
|
39 |
+
angles = get_3angle(rm_bkg_img, dino, val_preprocess, device)
|
40 |
|
41 |
phi = np.radians(angles[0])
|
42 |
theta = np.radians(angles[1])
|
43 |
gamma = angles[2]
|
44 |
+
confidence = float(angles[3])
|
45 |
+
if confidence > 0.5:
|
46 |
+
render_axis = render_3D_axis(phi, theta, gamma)
|
47 |
+
res_img = overlay_images_with_scaling(render_axis, rm_bkg_img)
|
48 |
+
else:
|
49 |
+
res_img = img
|
50 |
|
51 |
# axis_model = "axis.obj"
|
52 |
return [res_img, round(float(angles[0]), 2), round(float(angles[1]), 2), round(float(angles[2]), 2), round(float(angles[3]), 2)]
|
|
|
62 |
outputs=[
|
63 |
gr.Image(height=512, width=512, label="result image"),
|
64 |
# gr.Model3D(clear_color=[0.0, 0.0, 0.0, 0.0], label="3D Model"),
|
65 |
+
gr.Textbox(lines=1, label='Azimuth(0~360°) represents the position of the viewer in the xy plane'),
|
66 |
+
gr.Textbox(lines=1, label='Polar(-90~90°) indicating the height at which the viewer is located'),
|
67 |
+
gr.Textbox(lines=1, label='Rotation(-90~90°) represents the angle of rotation of the viewer'),
|
68 |
+
gr.Textbox(lines=1, label='Confidence(0~1) indicating whether the object has a meaningful orientation')
|
69 |
]
|
70 |
)
|
71 |
|
inference.py
ADDED
@@ -0,0 +1,49 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
from PIL import Image
|
3 |
+
from utils import *
|
4 |
+
import torch.nn.functional as F
|
5 |
+
import numpy as np
|
6 |
+
|
7 |
+
def get_3angle(image, dino, val_preprocess, device):
|
8 |
+
|
9 |
+
# image = Image.open(image_path).convert('RGB')
|
10 |
+
image_inputs = val_preprocess(images = image)
|
11 |
+
image_inputs['pixel_values'] = torch.from_numpy(np.array(image_inputs['pixel_values'])).to(device)
|
12 |
+
with torch.no_grad():
|
13 |
+
dino_pred = dino(image_inputs)
|
14 |
+
|
15 |
+
gaus_ax_pred = torch.argmax(dino_pred[:, 0:360], dim=-1)
|
16 |
+
gaus_pl_pred = torch.argmax(dino_pred[:, 360:360+180], dim=-1)
|
17 |
+
gaus_ro_pred = torch.argmax(dino_pred[:, 360+180:360+180+180], dim=-1)
|
18 |
+
confidence = F.softmax(dino_pred[:, -2:], dim=-1)[0][0]
|
19 |
+
angles = torch.zeros(4)
|
20 |
+
angles[0] = gaus_ax_pred
|
21 |
+
angles[1] = gaus_pl_pred - 90
|
22 |
+
angles[2] = gaus_ro_pred - 90
|
23 |
+
angles[3] = confidence
|
24 |
+
return angles
|
25 |
+
|
26 |
+
def get_3angle_infer_aug(origin_img, rm_bkg_img, dino, val_preprocess, device):
|
27 |
+
|
28 |
+
# image = Image.open(image_path).convert('RGB')
|
29 |
+
image = get_crop_images(origin_img, num=3) + get_crop_images(rm_bkg_img, num=3)
|
30 |
+
image_inputs = val_preprocess(images = image)
|
31 |
+
image_inputs['pixel_values'] = torch.from_numpy(np.array(image_inputs['pixel_values'])).to(device)
|
32 |
+
with torch.no_grad():
|
33 |
+
dino_pred = dino(image_inputs)
|
34 |
+
|
35 |
+
gaus_ax_pred = torch.argmax(dino_pred[:, 0:360], dim=-1).to(torch.float32)
|
36 |
+
gaus_pl_pred = torch.argmax(dino_pred[:, 360:360+180], dim=-1).to(torch.float32)
|
37 |
+
gaus_ro_pred = torch.argmax(dino_pred[:, 360+180:360+180+180], dim=-1).to(torch.float32)
|
38 |
+
|
39 |
+
gaus_ax_pred = remove_outliers_and_average_circular(gaus_ax_pred)
|
40 |
+
gaus_pl_pred = remove_outliers_and_average(gaus_pl_pred)
|
41 |
+
gaus_ro_pred = remove_outliers_and_average(gaus_ro_pred)
|
42 |
+
|
43 |
+
confidence = torch.mean(F.softmax(dino_pred[:, -2:], dim=-1), dim=0)[0]
|
44 |
+
angles = torch.zeros(4)
|
45 |
+
angles[0] = gaus_ax_pred
|
46 |
+
angles[1] = gaus_pl_pred - 90
|
47 |
+
angles[2] = gaus_ro_pred - 90
|
48 |
+
angles[3] = confidence
|
49 |
+
return angles
|