Spaces:
Running
on
Zero
Running
on
Zero
Update app.py
Browse files
app.py
CHANGED
@@ -7,176 +7,120 @@ from transformers import AutoImageProcessor, AutoModel
|
|
7 |
import torch.nn.functional as F
|
8 |
import spaces
|
9 |
|
10 |
-
|
11 |
DINO_MODELS = {
|
12 |
-
"DINOv3 Base ViT": "facebook/dinov3-vitb16-pretrain-lvd1689m",
|
13 |
"DINOv3 Large ViT": "facebook/dinov3-vitl16-pretrain-lvd1689m",
|
14 |
-
"DINOv3 Large ConvNeXT": "facebook/dinov3-convnext-large-pretrain-lvd1689m"
|
15 |
}
|
16 |
|
|
|
|
|
17 |
|
18 |
def load_model(model_name):
|
19 |
-
global processor
|
20 |
model_path = DINO_MODELS[model_name]
|
21 |
-
|
22 |
processor = AutoImageProcessor.from_pretrained(model_path)
|
23 |
-
model = AutoModel.from_pretrained(model_path)
|
24 |
return f"✅ Model '{model_name}' loaded successfully!"
|
25 |
|
26 |
-
load_model("DINOv3 Base ViT")
|
27 |
-
|
28 |
@spaces.GPU()
|
29 |
-
def extract_features(image):
|
30 |
-
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
31 |
|
32 |
-
|
|
|
|
|
|
|
|
|
|
|
33 |
original_size = image.size
|
34 |
-
inputs = processor(images=image, return_tensors="pt")
|
35 |
-
inputs = {k: v.to(device) for k, v in inputs.items()}
|
36 |
-
model_size = processor.size['height']
|
37 |
-
|
38 |
with torch.no_grad():
|
39 |
outputs = model(**inputs)
|
40 |
features = outputs.last_hidden_state
|
41 |
-
|
42 |
-
return features, original_size, model_size
|
43 |
|
44 |
def find_correspondences(features1, features2, threshold=0.8):
|
|
|
45 |
B, N1, D = features1.shape
|
46 |
-
|
47 |
-
|
48 |
features1_norm = F.normalize(features1, dim=-1)
|
49 |
features2_norm = F.normalize(features2, dim=-1)
|
50 |
-
|
51 |
similarity = torch.matmul(features1_norm, features2_norm.transpose(-2, -1))
|
52 |
-
|
53 |
matches1 = torch.argmax(similarity, dim=-1)
|
54 |
matches2 = torch.argmax(similarity, dim=-2)
|
55 |
-
|
56 |
max_sim1 = torch.max(similarity, dim=-1)[0]
|
57 |
-
|
58 |
-
|
59 |
-
mutual_matches = matches2[0, matches1[0]] == torch.arange(N1).to(device)
|
60 |
good_matches = (max_sim1[0] > threshold) & mutual_matches
|
61 |
-
|
62 |
-
return matches1[0][good_matches], torch.arange(N1).to(device)[good_matches], max_sim1[0][good_matches]
|
63 |
|
64 |
def patch_to_image_coords(patch_idx, original_size, model_size, patch_size=14):
|
65 |
orig_w, orig_h = original_size
|
66 |
patches_h = model_size // patch_size
|
67 |
patches_w = model_size // patch_size
|
68 |
-
|
69 |
if patch_idx >= patches_h * patches_w:
|
70 |
return None, None
|
71 |
-
|
72 |
patch_y = patch_idx // patches_w
|
73 |
patch_x = patch_idx % patches_w
|
74 |
-
|
75 |
y_model = patch_y * patch_size + patch_size // 2
|
76 |
x_model = patch_x * patch_size + patch_size // 2
|
77 |
-
|
78 |
x = int(x_model * orig_w / model_size)
|
79 |
y = int(y_model * orig_h / model_size)
|
80 |
-
|
81 |
return x, y
|
82 |
|
83 |
def match_keypoints(image1, image2, model_name):
|
84 |
if image1 is None or image2 is None:
|
85 |
return None
|
86 |
-
|
87 |
load_model(model_name)
|
88 |
-
|
89 |
-
|
90 |
-
|
91 |
-
|
92 |
-
features1, original_size1, model_size1 = extract_features(img1_pil)
|
93 |
-
features2, original_size2, model_size2 = extract_features(img2_pil)
|
94 |
-
|
95 |
-
features1 = features1[:, 1:, :]
|
96 |
-
features2 = features2[:, 1:, :]
|
97 |
-
|
98 |
matches2_idx, matches1_idx, similarities = find_correspondences(features1, features2, threshold=0.7)
|
99 |
-
|
100 |
img1_np = np.array(img1_pil)
|
101 |
img2_np = np.array(img2_pil)
|
102 |
-
|
103 |
h1, w1 = img1_np.shape[:2]
|
104 |
h2, w2 = img2_np.shape[:2]
|
105 |
-
|
106 |
result_img = np.zeros((max(h1, h2), w1 + w2, 3), dtype=np.uint8)
|
107 |
result_img[:h1, :w1] = img1_np
|
108 |
-
result_img[:h2, w1:w1+w2] = img2_np
|
109 |
-
|
110 |
-
|
111 |
-
|
112 |
-
keypoints2 = []
|
113 |
-
|
114 |
-
for i, (m1, m2, sim) in enumerate(zip(matches1_idx.cpu(), matches2_idx.cpu(), similarities.cpu())):
|
115 |
-
x1, y1 = patch_to_image_coords(m1.item(), original_size1, model_size1)
|
116 |
-
x2, y2 = patch_to_image_coords(m2.item(), original_size2, model_size2)
|
117 |
-
|
118 |
if x1 is not None and x2 is not None:
|
119 |
-
color = (np.random.randint(0, 255
|
120 |
-
|
121 |
-
|
122 |
-
|
123 |
-
|
124 |
-
cv2.circle(result_img, (x1, y1), 15, color, -1)
|
125 |
-
cv2.circle(result_img, (x2 + w1, y2), 15, color, -1)
|
126 |
-
cv2.line(result_img, (x1, y1), (x2 + w1, y2), color, 10)
|
127 |
-
|
128 |
-
|
129 |
-
|
130 |
return result_img
|
131 |
|
132 |
-
load_model("DINOv3 Base ViT")
|
133 |
-
|
134 |
with gr.Blocks(title="DINOv3 Keypoint Matching") as demo:
|
135 |
gr.Markdown("# DINOv3 For Keypoint Matching")
|
136 |
gr.Markdown("DINOv3 can be used to find matching features between two images.")
|
137 |
gr.Markdown("Upload two images to find corresponding keypoints using DINOv3 features, switch between different DINOv3 checkpoints.")
|
138 |
-
|
139 |
with gr.Row():
|
140 |
image1 = gr.Image(label="Image 1", type="numpy")
|
141 |
image2 = gr.Image(label="Image 2", type="numpy")
|
142 |
with gr.Column(scale=1):
|
143 |
-
|
144 |
model_selector = gr.Dropdown(
|
145 |
choices=list(DINO_MODELS.keys()),
|
146 |
-
value=
|
147 |
label="Select DINOv3 Model",
|
148 |
-
info="Choose the model size. Larger models may provide better features but require more memory."
|
149 |
)
|
150 |
-
|
151 |
-
# Add status bar
|
152 |
status_bar = gr.Textbox(
|
153 |
-
value="✅ Model '
|
154 |
label="Status",
|
155 |
interactive=False,
|
156 |
-
container=False
|
157 |
)
|
158 |
-
|
159 |
match_btn = gr.Button("Find Correspondences", variant="primary")
|
160 |
-
|
161 |
with gr.Column(scale=2):
|
162 |
output_image = gr.Image(label="Matched Keypoints")
|
163 |
-
|
164 |
-
|
165 |
-
fn=load_model,
|
166 |
-
inputs=[model_selector],
|
167 |
-
outputs=[status_bar]
|
168 |
-
)
|
169 |
-
|
170 |
-
match_btn.click(
|
171 |
-
fn=match_keypoints,
|
172 |
-
inputs=[image1, image2, model_selector],
|
173 |
-
outputs=[output_image]
|
174 |
-
)
|
175 |
-
|
176 |
gr.Examples(
|
177 |
examples=[["map.jpg", "street.jpg"], ["bee.JPG", "bee_edited.jpg"]],
|
178 |
-
inputs=[image1, image2]
|
179 |
)
|
180 |
|
181 |
if __name__ == "__main__":
|
182 |
-
demo.launch(
|
|
|
7 |
import torch.nn.functional as F
|
8 |
import spaces
|
9 |
|
|
|
10 |
DINO_MODELS = {
|
11 |
+
"DINOv3 Base ViT": "facebook/dinov3-vitb16-pretrain-lvd1689m",
|
12 |
"DINOv3 Large ViT": "facebook/dinov3-vitl16-pretrain-lvd1689m",
|
13 |
+
"DINOv3 Large ConvNeXT": "facebook/dinov3-convnext-large-pretrain-lvd1689m",
|
14 |
}
|
15 |
|
16 |
+
_default_model_name = "DINOv3 Base ViT"
|
17 |
+
processor = AutoImageProcessor.from_pretrained(DINO_MODELS[_default_model_name])
|
18 |
|
19 |
def load_model(model_name):
|
20 |
+
global processor
|
21 |
model_path = DINO_MODELS[model_name]
|
|
|
22 |
processor = AutoImageProcessor.from_pretrained(model_path)
|
|
|
23 |
return f"✅ Model '{model_name}' loaded successfully!"
|
24 |
|
|
|
|
|
25 |
@spaces.GPU()
|
26 |
+
def extract_features(image, model_name):
|
|
|
27 |
|
28 |
+
model_id = DINO_MODELS[model_name]
|
29 |
+
model = AutoModel.from_pretrained(model_id).to("cuda").eval()
|
30 |
+
local_processor = AutoImageProcessor.from_pretrained(model_id)
|
31 |
+
inputs = local_processor(images=image, return_tensors="pt")
|
32 |
+
inputs = {k: v.to("cuda") for k, v in inputs.items()}
|
33 |
+
model_size = local_processor.size["height"]
|
34 |
original_size = image.size
|
|
|
|
|
|
|
|
|
35 |
with torch.no_grad():
|
36 |
outputs = model(**inputs)
|
37 |
features = outputs.last_hidden_state
|
38 |
+
return features[:, 1:, :].float().cpu(), original_size, model_size
|
|
|
39 |
|
40 |
def find_correspondences(features1, features2, threshold=0.8):
|
41 |
+
device = torch.device("cpu")
|
42 |
B, N1, D = features1.shape
|
43 |
+
_, N2, _ = features2.shape
|
|
|
44 |
features1_norm = F.normalize(features1, dim=-1)
|
45 |
features2_norm = F.normalize(features2, dim=-1)
|
|
|
46 |
similarity = torch.matmul(features1_norm, features2_norm.transpose(-2, -1))
|
|
|
47 |
matches1 = torch.argmax(similarity, dim=-1)
|
48 |
matches2 = torch.argmax(similarity, dim=-2)
|
|
|
49 |
max_sim1 = torch.max(similarity, dim=-1)[0]
|
50 |
+
arange1 = torch.arange(N1, device=device)
|
51 |
+
mutual_matches = matches2[0, matches1[0]] == arange1
|
|
|
52 |
good_matches = (max_sim1[0] > threshold) & mutual_matches
|
53 |
+
return matches1[0][good_matches].cpu(), arange1[good_matches].cpu(), max_sim1[0][good_matches].cpu()
|
|
|
54 |
|
55 |
def patch_to_image_coords(patch_idx, original_size, model_size, patch_size=14):
|
56 |
orig_w, orig_h = original_size
|
57 |
patches_h = model_size // patch_size
|
58 |
patches_w = model_size // patch_size
|
|
|
59 |
if patch_idx >= patches_h * patches_w:
|
60 |
return None, None
|
|
|
61 |
patch_y = patch_idx // patches_w
|
62 |
patch_x = patch_idx % patches_w
|
|
|
63 |
y_model = patch_y * patch_size + patch_size // 2
|
64 |
x_model = patch_x * patch_size + patch_size // 2
|
|
|
65 |
x = int(x_model * orig_w / model_size)
|
66 |
y = int(y_model * orig_h / model_size)
|
|
|
67 |
return x, y
|
68 |
|
69 |
def match_keypoints(image1, image2, model_name):
|
70 |
if image1 is None or image2 is None:
|
71 |
return None
|
|
|
72 |
load_model(model_name)
|
73 |
+
img1_pil = Image.fromarray(image1).convert("RGB")
|
74 |
+
img2_pil = Image.fromarray(image2).convert("RGB")
|
75 |
+
features1, original_size1, model_size1 = extract_features(img1_pil, model_name)
|
76 |
+
features2, original_size2, model_size2 = extract_features(img2_pil, model_name)
|
|
|
|
|
|
|
|
|
|
|
|
|
77 |
matches2_idx, matches1_idx, similarities = find_correspondences(features1, features2, threshold=0.7)
|
|
|
78 |
img1_np = np.array(img1_pil)
|
79 |
img2_np = np.array(img2_pil)
|
|
|
80 |
h1, w1 = img1_np.shape[:2]
|
81 |
h2, w2 = img2_np.shape[:2]
|
|
|
82 |
result_img = np.zeros((max(h1, h2), w1 + w2, 3), dtype=np.uint8)
|
83 |
result_img[:h1, :w1] = img1_np
|
84 |
+
result_img[:h2, w1:w1 + w2] = img2_np
|
85 |
+
for m1, m2, _ in zip(matches1_idx, matches2_idx, similarities):
|
86 |
+
x1, y1 = patch_to_image_coords(int(m1), original_size1, model_size1)
|
87 |
+
x2, y2 = patch_to_image_coords(int(m2), original_size2, model_size2)
|
|
|
|
|
|
|
|
|
|
|
|
|
88 |
if x1 is not None and x2 is not None:
|
89 |
+
color = tuple(np.random.randint(0, 255, size=3).tolist())
|
90 |
+
cv2.circle(result_img, (x1, y1), 6, color, -1)
|
91 |
+
cv2.circle(result_img, (x2 + w1, y2), 6, color, -1)
|
92 |
+
cv2.line(result_img, (x1, y1), (x2 + w1, y2), color, 2)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
93 |
return result_img
|
94 |
|
|
|
|
|
95 |
with gr.Blocks(title="DINOv3 Keypoint Matching") as demo:
|
96 |
gr.Markdown("# DINOv3 For Keypoint Matching")
|
97 |
gr.Markdown("DINOv3 can be used to find matching features between two images.")
|
98 |
gr.Markdown("Upload two images to find corresponding keypoints using DINOv3 features, switch between different DINOv3 checkpoints.")
|
|
|
99 |
with gr.Row():
|
100 |
image1 = gr.Image(label="Image 1", type="numpy")
|
101 |
image2 = gr.Image(label="Image 2", type="numpy")
|
102 |
with gr.Column(scale=1):
|
|
|
103 |
model_selector = gr.Dropdown(
|
104 |
choices=list(DINO_MODELS.keys()),
|
105 |
+
value=_default_model_name,
|
106 |
label="Select DINOv3 Model",
|
107 |
+
info="Choose the model size. Larger models may provide better features but require more memory.",
|
108 |
)
|
|
|
|
|
109 |
status_bar = gr.Textbox(
|
110 |
+
value=f"✅ Model '{_default_model_name}' ready.",
|
111 |
label="Status",
|
112 |
interactive=False,
|
113 |
+
container=False,
|
114 |
)
|
|
|
115 |
match_btn = gr.Button("Find Correspondences", variant="primary")
|
|
|
116 |
with gr.Column(scale=2):
|
117 |
output_image = gr.Image(label="Matched Keypoints")
|
118 |
+
model_selector.change(fn=load_model, inputs=[model_selector], outputs=[status_bar])
|
119 |
+
match_btn.click(fn=match_keypoints, inputs=[image1, image2, model_selector], outputs=[output_image])
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
120 |
gr.Examples(
|
121 |
examples=[["map.jpg", "street.jpg"], ["bee.JPG", "bee_edited.jpg"]],
|
122 |
+
inputs=[image1, image2],
|
123 |
)
|
124 |
|
125 |
if __name__ == "__main__":
|
126 |
+
demo.launch()
|