File size: 6,691 Bytes
6cf6f4e
 
 
 
 
 
5840ebd
 
 
 
6cf6f4e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
import os

# gradio for visual demo
import gradio as gr

# transformers for easy access to nnet
os.system("pip install scipy")
os.system("pip install torch")
os.system("pip install scikit-learn")
os.system("pip install torchvision")

import numpy as np
import torch
import torchvision.transforms as transforms
from PIL  import ImageDraw, ImageColor, Image
from typing import Tuple
from scipy.ndimage import binary_closing, binary_opening
from sklearn.decomposition import PCA
from sklearn.neighbors import NearestNeighbors
from random import randint

### Models
standard_array = np.load('standard.npy')
pca_model = torch.hub.load("facebookresearch/dinov2", "dinov2_vitb14")
pca_model.eval()

### Parameters
IMAGENET_DEFAULT_MEAN = (0.485, 0.456, 0.406)
IMAGENET_DEFAULT_STD = (0.229, 0.224, 0.225)
    
smaller_edge_size = 448
interpolation_mode = transforms.InterpolationMode.BICUBIC
patch_size = pca_model.patch_size
background_threshold = 0.05

apply_opening = True
apply_closing = True
device = 'cpu'

def make_transform() -> transforms.Compose:
    return transforms.Compose([
        transforms.Resize(size=smaller_edge_size, interpolation=interpolation_mode, antialias=True),
        transforms.ToTensor(),
        transforms.Normalize(mean=IMAGENET_DEFAULT_MEAN, std=IMAGENET_DEFAULT_STD),
    ])

def prepare_image(image: Image) -> Tuple[torch.Tensor, Tuple[int, int]]:
    transform = make_transform()
    image_tensor = transform(image)
    resize_scale = image.width / image_tensor.shape[2]
    
    # Crop image to dimensions that are a multiple of the patch size
    height, width = image_tensor.shape[1:] # C x H x W
    cropped_width, cropped_height = width - width % patch_size, height - height % patch_size
    image_tensor = image_tensor[:, :cropped_height, :cropped_width]

    grid_size = (cropped_height // patch_size, cropped_width // patch_size) # h x w (TODO: check)
    return image_tensor, grid_size, resize_scale

def make_foreground_mask(tokens,
                         grid_size: Tuple[int, int]):
    projection = tokens @ standard_array
    mask = projection >= background_threshold
    mask = mask.reshape(*grid_size)
    if apply_opening:
        mask = binary_opening(mask)
    if apply_closing:
        mask = binary_closing(mask)
    return mask.flatten()

def render_patch_pca(image: Image, mask, filter_background, tokens, grid_size) -> Image:
    pca = PCA(n_components=3)
    if filter_background : pca.fit(tokens[mask])
    else : pca.fit(tokens)
    projected_tokens = pca.transform(tokens)

    t = torch.tensor(projected_tokens)
    t_min = t.min(dim=0, keepdim=True).values
    t_max = t.max(dim=0, keepdim=True).values
    normalized_t = (t - t_min) / (t_max - t_min)

    array = (normalized_t * 255).byte().numpy()
    if filter_background : array[~mask] = 0
    array = array.reshape(*grid_size, 3)

    return Image.fromarray(array).resize((image.width, image.height), 0)

def extract_features(img, image_tensor, filter_background, grid_size):
    with torch.inference_mode():
        image_batch = image_tensor.unsqueeze(0).to(device)
        tokens = pca_model.get_intermediate_layers(image_batch)[0].squeeze()
        mask = make_foreground_mask(tokens, grid_size)
        img_pca = render_patch_pca(img, mask, filter_background, tokens, grid_size)
    return tokens.cpu().numpy(), mask, img_pca
    
def compute_features(img, filter_background):
    image_tensor, grid_size, resize_scale = prepare_image(img)
    features, mask, img_pca = extract_features(img, image_tensor, filter_background, grid_size)
    return features, mask, grid_size, resize_scale, img_pca

def idx_to_source_position(idx, grid_size, resize_scale):
    row = (idx // grid_size[1])*pca_model.patch_size*resize_scale + pca_model.patch_size / 2
    col = (idx % grid_size[1])*pca_model.patch_size*resize_scale + pca_model.patch_size / 2
    return row, col

def compute_nn(features1, features2):
    knn = NearestNeighbors(n_neighbors=1)
    knn.fit(features1)
    distances, match2to1 = knn.kneighbors(features2)
    match2to1 = np.array(match2to1)
    return distances, match2to1

def compute_matches(img1, img2, lr_check, filter_background, display_matches_threshold):
    # compute features
    features1, mask1, grid_size1, resize_scale1, img_pca1 = compute_features(img1, filter_background)
    features2, mask2, grid_size2, resize_scale2, img_pca2 = compute_features(img2, filter_background)

    # match features
    distances2to1, match2to1 = compute_nn(features1, features2)
    distances1to2, match1to2 = compute_nn(features2, features1)
    
    # display matches
    draw1 = ImageDraw.Draw(img1)
    draw2 = ImageDraw.Draw(img2)
    
    if(img1.size[1] > img2.size[1]):
        img1 = img1.resize(img2.size)
        resize_scale1 = resize_scale2
    else:
        img2 = img2.resize(img1.size)
        resize_scale2 = resize_scale1
    
    offset = img1.size[0]
    merged_image = Image.new('RGB',(offset + img2.size[0], max(img1.size[1], img2.size[1])), (250,250,250))
    merged_image.paste(img1,(0,0))
    merged_image.paste(img2,(offset,0))
    draw = ImageDraw.Draw(merged_image)
    
    colormap = ImageColor.colormap
    for idx2, idx1 in enumerate(match2to1):
        if lr_check and match1to2[idx1] != idx2: continue        
        row1, col1 = idx_to_source_position(idx1, grid_size1, resize_scale1)
        row2, col2 = idx_to_source_position(idx2, grid_size2, resize_scale2)
        
        if filter_background and not mask1[idx1]: continue
        if filter_background and not mask2[idx2]: continue
        
        r = randint(0,255)
        g = randint(0,255)
        color = (r,g,255-r)
        
        draw1.point((col1, row1), color)
        draw2.point((col2, row2), color)
        
        if 100*np.random.rand() > display_matches_threshold: continue 
        draw.line((col1, row1, col2 + offset, row2), fill=color)
        
    return [[img1, img_pca1], [img2, img_pca2], merged_image]

iface = gr.Interface(fn=compute_matches, 
    inputs=[
        gr.Image(type="pil"), 
        gr.Image(type="pil"), 
        gr.Checkbox(label="Keep only symmetric matches",),
        gr.Checkbox(label="Mask background"),
        gr.Slider(0, 100, step=5, value=5, label="Display matches ratio", info="Choose between 0 and 100%"),
    ], 
    outputs=[
        gr.Gallery(
            label="Image 1", show_label=False, elem_id="gallery",
            columns=[2], rows=[1], object_fit="contain", height="auto"), 
        gr.Gallery(
            label="Image 1", show_label=False, elem_id="gallery",
            columns=[2], rows=[1], object_fit="contain", height="auto"), 
        gr.Image(type="pil")
    ])
iface.launch(debug=True)