Spaces:
Running
Running
Martin Tomov
commited on
Commit
β’
850cda3
1
Parent(s):
69ed2a1
gsl_utils.py
Browse files- gsl_utils.py +122 -0
gsl_utils.py
ADDED
@@ -0,0 +1,122 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# GSL
|
2 |
+
|
3 |
+
import os
|
4 |
+
import torch
|
5 |
+
import numpy as np
|
6 |
+
from PIL import Image, ImageChops, ImageEnhance
|
7 |
+
import cv2
|
8 |
+
from simple_lama_inpainting import SimpleLama
|
9 |
+
from segment_anything import build_sam, SamPredictor
|
10 |
+
from GroundingDINO.groundingdino.util import box_ops
|
11 |
+
from GroundingDINO.groundingdino.util.slconfig import SLConfig
|
12 |
+
from GroundingDINO.groundingdino.util.utils import clean_state_dict
|
13 |
+
from GroundingDINO.groundingdino.util.inference import annotate, load_image, predict
|
14 |
+
from huggingface_hub import hf_hub_download
|
15 |
+
|
16 |
+
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
|
17 |
+
|
18 |
+
def load_model_hf(repo_id, filename, ckpt_config_filename, device='cpu'):
|
19 |
+
cache_config_file = hf_hub_download(repo_id=repo_id, filename=ckpt_config_filename)
|
20 |
+
args = SLConfig.fromfile(cache_config_file)
|
21 |
+
args.device = device
|
22 |
+
model = build_model(args)
|
23 |
+
cache_file = hf_hub_download(repo_id=repo_id, filename=filename)
|
24 |
+
checkpoint = torch.load(cache_file, map_location=device)
|
25 |
+
model.load_state_dict(clean_state_dict(checkpoint['model']), strict=False)
|
26 |
+
model.eval()
|
27 |
+
return model
|
28 |
+
|
29 |
+
groundingdino_model = load_model_hf(
|
30 |
+
repo_id="ShilongLiu/GroundingDINO",
|
31 |
+
filename="groundingdino_swinb_cogcoor.pth",
|
32 |
+
ckpt_config_filename="GroundingDINO_SwinB.cfg.py",
|
33 |
+
device=device
|
34 |
+
)
|
35 |
+
|
36 |
+
sam_predictor = SamPredictor(build_sam(checkpoint='sam_vit_h_4b8939.pth').to(device))
|
37 |
+
simple_lama = SimpleLama()
|
38 |
+
|
39 |
+
def detect(image, model, text_prompt='insect . flower . cloud', box_threshold=0.15, text_threshold=0.15):
|
40 |
+
boxes, logits, phrases = predict(
|
41 |
+
image=image,
|
42 |
+
model=model,
|
43 |
+
caption=text_prompt,
|
44 |
+
box_threshold=box_threshold,
|
45 |
+
text_threshold=text_threshold
|
46 |
+
)
|
47 |
+
annotated_frame = annotate(image_source=image, boxes=boxes, logits=logits, phrases=phrases)
|
48 |
+
annotated_frame = annotated_frame[..., ::-1] # BGR to RGB
|
49 |
+
return annotated_frame, boxes, phrases
|
50 |
+
|
51 |
+
def segment(image, sam_model, boxes):
|
52 |
+
sam_model.set_image(image)
|
53 |
+
H, W, _ = image.shape
|
54 |
+
boxes_xyxy = box_ops.box_cxcywh_to_xyxy(boxes) * torch.Tensor([W, H, W, H])
|
55 |
+
transformed_boxes = sam_model.transform.apply_boxes_torch(boxes_xyxy.to(device), image.shape[:2])
|
56 |
+
masks, _, _ = sam_model.predict_torch(
|
57 |
+
point_coords=None,
|
58 |
+
point_labels=None,
|
59 |
+
boxes=transformed_boxes,
|
60 |
+
multimask_output=True,
|
61 |
+
)
|
62 |
+
return masks.cpu()
|
63 |
+
|
64 |
+
def draw_mask(mask, image, random_color=True):
|
65 |
+
if random_color:
|
66 |
+
color = np.concatenate([np.random.random(3), np.array([0.8])], axis=0)
|
67 |
+
else:
|
68 |
+
color = np.array([30 / 255, 144 / 255, 255 / 255, 0.6])
|
69 |
+
h, w = mask.shape[-2:]
|
70 |
+
mask_image = mask.reshape(h, w, 1) * color.reshape(1, 1, -1)
|
71 |
+
annotated_frame_pil = Image.fromarray(image).convert("RGBA")
|
72 |
+
mask_image_pil = Image.fromarray((mask_image.numpy() * 255).astype(np.uint8)).convert("RGBA")
|
73 |
+
return np.array(Image.alpha_composite(annotated_frame_pil, mask_image_pil))
|
74 |
+
|
75 |
+
def dilate_mask(mask, dilate_factor=15):
|
76 |
+
mask = mask.astype(np.uint8)
|
77 |
+
mask = cv2.dilate(
|
78 |
+
mask,
|
79 |
+
np.ones((dilate_factor, dilate_factor), np.uint8),
|
80 |
+
iterations=1
|
81 |
+
)
|
82 |
+
return mask
|
83 |
+
|
84 |
+
def gsl_process_image(local_image_path):
|
85 |
+
# Load image
|
86 |
+
image_source, image = load_image(local_image_path)
|
87 |
+
|
88 |
+
# Detect insects
|
89 |
+
annotated_frame, detected_boxes, phrases = detect(image, model=groundingdino_model)
|
90 |
+
indices = [i for i, s in enumerate(phrases) if 'insect' in s]
|
91 |
+
|
92 |
+
# Segment insects
|
93 |
+
segmented_frame_masks = segment(image_source, sam_predictor, detected_boxes[indices])
|
94 |
+
|
95 |
+
# Combine masks
|
96 |
+
final_mask = None
|
97 |
+
for i in range(len(segmented_frame_masks) - 1):
|
98 |
+
if final_mask is None:
|
99 |
+
final_mask = np.bitwise_or(segmented_frame_masks[i][0].cpu(), segmented_frame_masks[i + 1][0].cpu())
|
100 |
+
else:
|
101 |
+
final_mask = np.bitwise_or(final_mask, segmented_frame_masks[i + 1][0].cpu())
|
102 |
+
|
103 |
+
# Draw mask
|
104 |
+
annotated_frame_with_mask = draw_mask(final_mask, image_source)
|
105 |
+
|
106 |
+
# Dilate mask
|
107 |
+
mask = final_mask.numpy()
|
108 |
+
mask = mask.astype(np.uint8) * 255
|
109 |
+
mask = dilate_mask(mask)
|
110 |
+
dilated_image_mask_pil = Image.fromarray(mask)
|
111 |
+
|
112 |
+
# Inpainting
|
113 |
+
result = simple_lama(image_source, dilated_image_mask_pil)
|
114 |
+
|
115 |
+
# Difference and composite
|
116 |
+
diff = ImageChops.difference(result, Image.fromarray(image_source))
|
117 |
+
threshold = 7
|
118 |
+
diff2 = diff.convert('L').point(lambda p: 255 if p > threshold else 0).convert('1')
|
119 |
+
img3 = Image.new('RGB', Image.fromarray(image_source).size, (255, 236, 10))
|
120 |
+
diff3 = Image.composite(Image.fromarray(image_source), img3, diff2)
|
121 |
+
|
122 |
+
return diff3
|