haritsahm
commited on
Commit
·
fb345ee
1
Parent(s):
f0c912f
Add box selection feature
Browse files- app.py +91 -11
- utils/utils.py +23 -0
app.py
CHANGED
@@ -13,6 +13,77 @@ from utils import utils
|
|
13 |
SAM_MODEL = utils.get_model('vit_b')
|
14 |
|
15 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
16 |
def click_process(model, show_mask, radius_width):
|
17 |
|
18 |
bg_image = st.session_state['image']
|
@@ -24,7 +95,7 @@ def click_process(model, show_mask, radius_width):
|
|
24 |
if 'result_image' not in st.session_state:
|
25 |
st.session_state.result_image = bg_image.resize(scaled_hw)
|
26 |
|
27 |
-
|
28 |
fill_color="rgba(255, 255, 0, 0.8)",
|
29 |
background_image = bg_image,
|
30 |
drawing_mode='point',
|
@@ -35,13 +106,21 @@ def click_process(model, show_mask, radius_width):
|
|
35 |
update_streamlit=True,
|
36 |
key="point",)
|
37 |
|
38 |
-
# ! Warn: Can cause infinite loop or high cpu usage
|
39 |
if not show_mask:
|
40 |
-
|
41 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
42 |
|
43 |
-
elif
|
44 |
-
df = pd.json_normalize(
|
45 |
input_points = []
|
46 |
input_labels = []
|
47 |
|
@@ -88,6 +167,8 @@ def image_preprocess_callback(model):
|
|
88 |
st.session_state.image = image
|
89 |
else:
|
90 |
with st.spinner(text="Cleaning up!"):
|
|
|
|
|
91 |
if 'image' in st.session_state:
|
92 |
st.session_state.image = None
|
93 |
if 'result_image' in st.session_state:
|
@@ -130,16 +211,15 @@ def main():
|
|
130 |
if option == 'Click':
|
131 |
with st.spinner(text="Computing masks"):
|
132 |
result_image = click_process(SAM_MODEL, show_mask, radius_width)
|
|
|
|
|
133 |
with canvas_output:
|
134 |
if result_image is not None:
|
135 |
st.write("Result")
|
136 |
st.image(result_image)
|
137 |
|
138 |
-
|
139 |
-
|
140 |
-
# if 'image' in st.session_state:
|
141 |
-
# if st.session_state.image is None:
|
142 |
-
# st.session_state.clear()
|
143 |
|
144 |
|
145 |
if __name__ == '__main__':
|
|
|
13 |
SAM_MODEL = utils.get_model('vit_b')
|
14 |
|
15 |
|
16 |
+
def box_process(model, show_mask, radius_width):
|
17 |
+
bg_image = st.session_state['image']
|
18 |
+
width, height = bg_image.size[:2]
|
19 |
+
container_width = 700
|
20 |
+
scale = container_width/width
|
21 |
+
scaled_hw = (container_width, int(height * scale))
|
22 |
+
|
23 |
+
if 'result_image' not in st.session_state:
|
24 |
+
st.session_state.result_image = bg_image.resize(scaled_hw)
|
25 |
+
|
26 |
+
box_canvas = st_canvas(
|
27 |
+
fill_color="rgba(255, 255, 0, 0)",
|
28 |
+
background_image = bg_image,
|
29 |
+
drawing_mode='rect',
|
30 |
+
stroke_color = "rgba(0, 255, 0, 0.6)",
|
31 |
+
stroke_width = radius_width,
|
32 |
+
width = container_width,
|
33 |
+
height = height * scale,
|
34 |
+
point_display_radius = 12,
|
35 |
+
update_streamlit=True,
|
36 |
+
key="box"
|
37 |
+
)
|
38 |
+
|
39 |
+
if not show_mask:
|
40 |
+
if 'rerun_once' in st.session_state:
|
41 |
+
if st.session_state.rerun_once:
|
42 |
+
st.session_state.rerun_once = False
|
43 |
+
else:
|
44 |
+
st.session_state.rerun_once = True
|
45 |
+
|
46 |
+
st.session_state.display_result = True
|
47 |
+
if st.session_state.rerun_once:
|
48 |
+
st.experimental_rerun()
|
49 |
+
else:
|
50 |
+
return np.asarray(bg_image)
|
51 |
+
|
52 |
+
elif box_canvas.json_data is not None:
|
53 |
+
df = pd.json_normalize(box_canvas.json_data["objects"])
|
54 |
+
center_point,center_label,input_box = [],[],[]
|
55 |
+
center_point, center_label, input_box = [], [], []
|
56 |
+
for _, row in df.iterrows():
|
57 |
+
x, y, w,h = row["left"], row["top"], row["width"], row["height"]
|
58 |
+
x = int(x/scale)
|
59 |
+
y = int(y/scale)
|
60 |
+
w = int(w/scale)
|
61 |
+
h = int(h/scale)
|
62 |
+
center_point.append([x+w/2,y+h/2])
|
63 |
+
center_label.append([1])
|
64 |
+
input_box.append([x,y,x+w,y+h])
|
65 |
+
|
66 |
+
masks = []
|
67 |
+
if model:
|
68 |
+
masks = utils.model_predict_masks_box(model, center_point, center_label, input_box)
|
69 |
+
|
70 |
+
if len(masks) == 0:
|
71 |
+
return bg_image
|
72 |
+
|
73 |
+
bg_image = np.asarray(bg_image)
|
74 |
+
color = np.concatenate([random.choice(utils.get_color()), np.array([0.6])], axis=0)
|
75 |
+
im_masked = utils.show_click(masks,color)
|
76 |
+
im_masked = Image.fromarray(im_masked).convert('RGBA')
|
77 |
+
result_image = Image.alpha_composite(Image.fromarray(bg_image).convert('RGBA'),im_masked).convert("RGB")
|
78 |
+
result_image = result_image.resize(scaled_hw)
|
79 |
+
st.session_state.display_result = True
|
80 |
+
return result_image
|
81 |
+
else:
|
82 |
+
return np.asarray(bg_image)
|
83 |
+
|
84 |
+
return np.asarray(bg_image)
|
85 |
+
|
86 |
+
|
87 |
def click_process(model, show_mask, radius_width):
|
88 |
|
89 |
bg_image = st.session_state['image']
|
|
|
95 |
if 'result_image' not in st.session_state:
|
96 |
st.session_state.result_image = bg_image.resize(scaled_hw)
|
97 |
|
98 |
+
click_canvas = st_canvas(
|
99 |
fill_color="rgba(255, 255, 0, 0.8)",
|
100 |
background_image = bg_image,
|
101 |
drawing_mode='point',
|
|
|
106 |
update_streamlit=True,
|
107 |
key="point",)
|
108 |
|
|
|
109 |
if not show_mask:
|
110 |
+
if 'rerun_once' in st.session_state:
|
111 |
+
if st.session_state.rerun_once:
|
112 |
+
st.session_state.rerun_once = False
|
113 |
+
else:
|
114 |
+
st.session_state.rerun_once = True
|
115 |
+
|
116 |
+
st.session_state.display_result = True
|
117 |
+
if st.session_state.rerun_once:
|
118 |
+
st.experimental_rerun()
|
119 |
+
else:
|
120 |
+
return np.asarray(bg_image)
|
121 |
|
122 |
+
elif click_canvas.json_data is not None:
|
123 |
+
df = pd.json_normalize(click_canvas.json_data["objects"])
|
124 |
input_points = []
|
125 |
input_labels = []
|
126 |
|
|
|
167 |
st.session_state.image = image
|
168 |
else:
|
169 |
with st.spinner(text="Cleaning up!"):
|
170 |
+
if 'display_result' in st.session_state:
|
171 |
+
st.session_state.display_result = False
|
172 |
if 'image' in st.session_state:
|
173 |
st.session_state.image = None
|
174 |
if 'result_image' in st.session_state:
|
|
|
211 |
if option == 'Click':
|
212 |
with st.spinner(text="Computing masks"):
|
213 |
result_image = click_process(SAM_MODEL, show_mask, radius_width)
|
214 |
+
elif option == 'Box':
|
215 |
+
result_image = box_process(SAM_MODEL, show_mask, radius_width)
|
216 |
with canvas_output:
|
217 |
if result_image is not None:
|
218 |
st.write("Result")
|
219 |
st.image(result_image)
|
220 |
|
221 |
+
else:
|
222 |
+
st.cache_data.clear()
|
|
|
|
|
|
|
223 |
|
224 |
|
225 |
if __name__ == '__main__':
|
utils/utils.py
CHANGED
@@ -70,3 +70,26 @@ def model_predict_masks_click(model,input_points,input_labels):
|
|
70 |
torch.cuda.empty_cache()
|
71 |
|
72 |
return masks
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
70 |
torch.cuda.empty_cache()
|
71 |
|
72 |
return masks
|
73 |
+
|
74 |
+
def model_predict_masks_box(model,center_point,center_label,input_box):
|
75 |
+
masks = np.array([])
|
76 |
+
for i in range(len(center_label)):
|
77 |
+
if center_point[i] == []:continue
|
78 |
+
center_point_1 = np.array([center_point[i]])
|
79 |
+
center_label_1 = np.array(center_label[i])
|
80 |
+
input_box_1 = np.array(input_box[i])
|
81 |
+
mask, _, _ = model.predict(
|
82 |
+
point_coords=center_point_1,
|
83 |
+
point_labels=center_label_1,
|
84 |
+
box=input_box_1,
|
85 |
+
multimask_output=False,
|
86 |
+
)
|
87 |
+
try:
|
88 |
+
masks = masks + mask
|
89 |
+
except:
|
90 |
+
masks = mask
|
91 |
+
|
92 |
+
if torch.cuda.is_available():
|
93 |
+
torch.cuda.empty_cache()
|
94 |
+
|
95 |
+
return masks
|