alexanderkroner
commited on
Update README.md
Browse files
README.md
CHANGED
@@ -21,6 +21,145 @@ MSI-Net is a visual saliency model that predicts where humans fixate on natural
|
|
21 |
|
22 |
<img src="https://github.com/alexanderkroner/saliency/blob/master/figures/architecture.jpg?raw=true" width="700">
|
23 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
24 |
# Datasets
|
25 |
|
26 |
Before training the model on fixation data, the encoder weights were initialized from a VGG16 backbone pre-trained on the ImageNet classification task. The model was then trained on the SALICON dataset, which consists of mouse movement recordings as a proxy for gaze measurements. Finally, the weights can be fine-tuned on human eye tracking data. MSI-Net was therefore also trained on one of the following datasets, although here we only provide the SALICON base model:
|
|
|
21 |
|
22 |
<img src="https://github.com/alexanderkroner/saliency/blob/master/figures/architecture.jpg?raw=true" width="700">
|
23 |
|
24 |
+
# Example Use
|
25 |
+
|
26 |
+
### Import the dependencies
|
27 |
+
```python
|
28 |
+
import matplotlib.pyplot as plt
|
29 |
+
import numpy as np
|
30 |
+
import tensorflow as tf
|
31 |
+
from huggingface_hub import snapshot_download
|
32 |
+
```
|
33 |
+
|
34 |
+
### Download the repo files
|
35 |
+
```python
|
36 |
+
hf_dir = snapshot_download(
|
37 |
+
repo_id="alexanderkroner/MSI-Net", allow_patterns=["*.pb", "*.jpg"]
|
38 |
+
)
|
39 |
+
```
|
40 |
+
|
41 |
+
### Load the saliency model
|
42 |
+
```python
|
43 |
+
model = tf.saved_model.load(hf_dir)
|
44 |
+
model = model.signatures["serving_default"]
|
45 |
+
```
|
46 |
+
|
47 |
+
### Load the functions for preprocessing the input and postprocessing the output
|
48 |
+
```python
|
49 |
+
def get_target_shape(original_shape):
|
50 |
+
original_aspect_ratio = original_shape[0] / original_shape[1]
|
51 |
+
|
52 |
+
square_mode = abs(original_aspect_ratio - 1.0)
|
53 |
+
landscape_mode = abs(original_aspect_ratio - 240 / 320)
|
54 |
+
portrait_mode = abs(original_aspect_ratio - 320 / 240)
|
55 |
+
|
56 |
+
best_mode = min(square_mode, landscape_mode, portrait_mode)
|
57 |
+
|
58 |
+
if best_mode == square_mode:
|
59 |
+
target_shape = (320, 320)
|
60 |
+
elif best_mode == landscape_mode:
|
61 |
+
target_shape = (240, 320)
|
62 |
+
else:
|
63 |
+
target_shape = (320, 240)
|
64 |
+
|
65 |
+
return target_shape
|
66 |
+
|
67 |
+
|
68 |
+
def preprocess_input(input_image, target_shape):
|
69 |
+
input_tensor = tf.expand_dims(input_image, axis=0)
|
70 |
+
|
71 |
+
input_tensor = tf.image.resize(
|
72 |
+
input_tensor, target_shape, preserve_aspect_ratio=True
|
73 |
+
)
|
74 |
+
|
75 |
+
vertical_padding = target_shape[0] - input_tensor.shape[1]
|
76 |
+
horizontal_padding = target_shape[1] - input_tensor.shape[2]
|
77 |
+
|
78 |
+
vertical_padding_1 = vertical_padding // 2
|
79 |
+
vertical_padding_2 = vertical_padding - vertical_padding_1
|
80 |
+
|
81 |
+
horizontal_padding_1 = horizontal_padding // 2
|
82 |
+
horizontal_padding_2 = horizontal_padding - horizontal_padding_1
|
83 |
+
|
84 |
+
input_tensor = tf.pad(
|
85 |
+
input_tensor,
|
86 |
+
[
|
87 |
+
[0, 0],
|
88 |
+
[vertical_padding_1, vertical_padding_2],
|
89 |
+
[horizontal_padding_1, horizontal_padding_2],
|
90 |
+
[0, 0],
|
91 |
+
],
|
92 |
+
)
|
93 |
+
|
94 |
+
return (
|
95 |
+
input_tensor,
|
96 |
+
[vertical_padding_1, vertical_padding_2],
|
97 |
+
[horizontal_padding_1, horizontal_padding_2],
|
98 |
+
)
|
99 |
+
|
100 |
+
|
101 |
+
def postprocess_output(
|
102 |
+
output_tensor, vertical_padding, horizontal_padding, original_shape
|
103 |
+
):
|
104 |
+
output_tensor = output_tensor[
|
105 |
+
:,
|
106 |
+
vertical_padding[0] : output_tensor.shape[1] - vertical_padding[1],
|
107 |
+
horizontal_padding[0] : output_tensor.shape[2] - horizontal_padding[1],
|
108 |
+
:,
|
109 |
+
]
|
110 |
+
|
111 |
+
output_tensor = tf.image.resize(output_tensor, original_shape)
|
112 |
+
|
113 |
+
output_array = output_tensor.numpy().squeeze()
|
114 |
+
output_array = plt.cm.inferno(output_array)[..., :3]
|
115 |
+
|
116 |
+
return output_array
|
117 |
+
```
|
118 |
+
|
119 |
+
### Load and preprocess an example image
|
120 |
+
```python
|
121 |
+
input_image = tf.keras.utils.load_img(hf_dir + "/example.jpg")
|
122 |
+
input_image = np.array(input_image, dtype=np.float32)
|
123 |
+
|
124 |
+
original_shape = input_image.shape[:2]
|
125 |
+
target_shape = get_target_shape(original_shape)
|
126 |
+
|
127 |
+
input_tensor, vertical_padding, horizontal_padding = preprocess_input(
|
128 |
+
input_image, target_shape
|
129 |
+
)
|
130 |
+
```
|
131 |
+
|
132 |
+
### Feed the input tensor to the model
|
133 |
+
```python
|
134 |
+
output_tensor = model(input_tensor)["output"]
|
135 |
+
```
|
136 |
+
|
137 |
+
### Postprocess and visualize the output
|
138 |
+
```python
|
139 |
+
saliency_map = postprocess_output(
|
140 |
+
output_tensor, vertical_padding, horizontal_padding, original_shape
|
141 |
+
)
|
142 |
+
|
143 |
+
alpha = 0.65
|
144 |
+
|
145 |
+
blended_image = alpha * saliency_map + (1 - alpha) * input_image / 255
|
146 |
+
|
147 |
+
plt.figure(figsize=(10, 5))
|
148 |
+
|
149 |
+
plt.subplot(1, 2, 1)
|
150 |
+
plt.imshow(input_image / 255)
|
151 |
+
plt.title("Input Image")
|
152 |
+
plt.axis("off")
|
153 |
+
|
154 |
+
plt.subplot(1, 2, 2)
|
155 |
+
plt.imshow(blended_image)
|
156 |
+
plt.title("Saliency Map")
|
157 |
+
plt.axis("off")
|
158 |
+
|
159 |
+
plt.tight_layout()
|
160 |
+
plt.show()
|
161 |
+
```
|
162 |
+
|
163 |
# Datasets
|
164 |
|
165 |
Before training the model on fixation data, the encoder weights were initialized from a VGG16 backbone pre-trained on the ImageNet classification task. The model was then trained on the SALICON dataset, which consists of mouse movement recordings as a proxy for gaze measurements. Finally, the weights can be fine-tuned on human eye tracking data. MSI-Net was therefore also trained on one of the following datasets, although here we only provide the SALICON base model:
|