Skip to content

Commit 813ec15

Browse files
AllakhverdovAllakhverdov
authored
Merge pull request #2 from mrsndmn/main
Project page
2 parents a7122ef + eb196c3 commit 813ec15

27 files changed

+3289
-0
lines changed

‎.github/workflows/project-page.yaml

Lines changed: 43 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,43 @@
1+
name: Deploy Project Page
2+
3+
on:
4+
push:
5+
branches:
6+
- main
7+
paths:
8+
- project/**
9+
10+
jobs:
11+
build-project-page:
12+
name: 'Project page static files artifact'
13+
runs-on: ubuntu-latest
14+
15+
steps:
16+
- name: Check out code
17+
uses: actions/checkout@v4
18+
19+
- name: Copy project page static files
20+
uses: actions/upload-pages-artifact@v3
21+
with:
22+
path: ./project
23+
24+
deploy-project-page:
25+
# Add a dependency to the build job
26+
needs: build-project-page
27+
28+
# Grant GITHUB_TOKEN the permissions required to make a Pages deployment
29+
permissions:
30+
pages: write # to deploy to Pages
31+
id-token: write # to verify the deployment originates from an appropriate source
32+
33+
# Deploy to the github-pages environment
34+
environment:
35+
name: github-pages
36+
url: ${{ steps.deployment.outputs.page_url }}
37+
38+
# Specify runner + deployment step
39+
runs-on: ubuntu-latest
40+
steps:
41+
- name: Deploy to GitHub Pages
42+
id: deployment
43+
uses: actions/deploy-pages@v4 # or specific "vX.X.X" version tag for this action

‎project/README.md

Lines changed: 3 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,3 @@
1+
[**Academic Project Page Template**](https://github.com/eliahuhorwitz/Academic-project-page-template.git) used for project page
2+
3+

‎project/index.html

Lines changed: 237 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,237 @@
1+
<!DOCTYPE html>
2+
<html>
3+
4+
<head>
5+
<meta charset="utf-8">
6+
<!-- Meta tags for social media banners, these should be filled in appropriatly as they are your "business card" -->
7+
<!-- Replace the content tag with appropriate information -->
8+
<meta name="description" content="Image Reconstruction as a Tool for Feature Analysis">
9+
<meta property="og:title" content="Image Reconstruction as a Tool for Feature Analysis" />
10+
<meta property="og:description"
11+
content="A novel approach for interpreting vision features via image reconstruction" />
12+
<meta property="og:url" content="https://fusionbrainlab.github.io/feature_analysis" />
13+
<!-- Path to banner image, should be in the path listed below. Optimal dimenssions are 1200X630-->
14+
<meta property="og:image" content="static/images/v1_vs_v2.png" />
15+
<meta property="og:image:width" content="1200" />
16+
<meta property="og:image:height" content="630" />
17+
18+
19+
<meta name="twitter:title" content="Image Reconstruction as a Tool for Feature Analysis">
20+
<meta name="twitter:description" content="A novel approach for interpreting vision features via image reconstruction">
21+
<!-- Path to banner image, should be in the path listed below. Optimal dimenssions are 1200X600-->
22+
<meta name="twitter:image" content="static/images/v1_vs_v2.png">
23+
<meta name="twitter:card" content="summary_large_image">
24+
<!-- Keywords for your paper to be indexed by-->
25+
<meta name="keywords" content="computer vision, feature analysis, image reconstruction, vision encoders">
26+
<meta name="viewport" content="width=device-width, initial-scale=1">
27+
28+
29+
<title>Image Reconstruction as a Tool for Feature Analysis</title>
30+
<link rel="icon" type="image/x-icon" href="static/images/favicon.ico">
31+
<link href="https://fonts.googleapis.com/css?family=Google+Sans|Noto+Sans|Castoro" rel="stylesheet">
32+
33+
<link rel="stylesheet" href="static/css/bulma.min.css">
34+
<link rel="stylesheet" href="static/css/bulma-carousel.min.css">
35+
<link rel="stylesheet" href="static/css/bulma-slider.min.css">
36+
<link rel="stylesheet" href="static/css/fontawesome.all.min.css">
37+
<link rel="stylesheet" href="https://cdn.jsdelivr.net/gh/jpswalsh/academicons@1/css/academicons.min.css">
38+
<link rel="stylesheet" href="static/css/index.css">
39+
40+
<script src="https://ajax.googleapis.com/ajax/libs/jquery/3.5.1/jquery.min.js"></script>
41+
<script src="https://documentcloud.adobe.com/view-sdk/main.js"></script>
42+
<script defer src="static/js/fontawesome.all.min.js"></script>
43+
<script src="static/js/bulma-carousel.min.js"></script>
44+
<script src="static/js/bulma-slider.min.js"></script>
45+
<script src="static/js/index.js"></script>
46+
</head>
47+
48+
<body>
49+
50+
51+
<section class="hero">
52+
<div class="hero-body">
53+
<div class="container is-max-desktop">
54+
<div class="columns is-centered">
55+
<div class="column has-text-centered">
56+
<h1 class="title is-1 publication-title">Image Reconstruction as a Tool for Feature Analysis</h1>
57+
<div class="is-size-5 publication-authors">
58+
<!-- Paper authors -->
59+
<span class="author-block">
60+
<a href="mailto:[email protected]" target="_blank">Eduard Allakhverdov</a>
61+
</span>
62+
<span class="author-block">
63+
<a href="mailto:[email protected]" target="_blank">Dmitrii Tarasov</a>
64+
</span>
65+
<span class="author-block">
66+
<a href="mailto:[email protected]" target="_blank">Elizaveta Goncharova</a>
67+
</span>
68+
<span class="author-block">
69+
<a href="mailto:[email protected]" target="_blank">Andrey Kuznetsov</a>
70+
</span>
71+
</div>
72+
<div class="is-size-5 publication-authors">
73+
<span class="author-block">
74+
AIRI, Moscow, Russia<br>
75+
MIPT, Dolgoprudny, Russia
76+
</span>
77+
<span class="author-block">
78+
AIRI, Moscow, Russia
79+
</span>
80+
<span class="author-block">
81+
AIRI, Moscow, Russia
82+
</span>
83+
<span class="author-block">
84+
AIRI, Moscow, Russia
85+
</span>
86+
</div>
87+
88+
<div class="column has-text-centered">
89+
<div class="publication-links">
90+
91+
<!-- Github link -->
92+
<span class="link-block">
93+
<a href="https://github.com/FusionBrainLab/feature_analysis" target="_blank"
94+
class="external-link button is-normal is-rounded is-dark">
95+
<span class="icon">
96+
<i class="fab fa-github"></i>
97+
</span>
98+
<span>Code</span>
99+
</a>
100+
</span>
101+
102+
<!-- ArXiv abstract Link -->
103+
<span class="link-block">
104+
<a href="https://arxiv.org/abs/<ARXIV PAPER ID>" target="_blank"
105+
class="external-link button is-normal is-rounded is-dark">
106+
<span class="icon">
107+
<i class="ai ai-arxiv"></i>
108+
</span>
109+
<span>arXiv</span>
110+
</a>
111+
</span>
112+
</div>
113+
</div>
114+
</div>
115+
</div>
116+
</section>
117+
118+
119+
120+
<!-- Paper abstract -->
121+
<section class="section hero is-light">
122+
<div class="container is-max-desktop">
123+
<div class="columns is-centered has-text-centered">
124+
<div class="column is-four-fifths">
125+
<h2 class="title is-3">Abstract</h2>
126+
<div class="content has-text-justified">
127+
<p>
128+
Vision encoders are increasingly used in modern applications, from vision-only models to multimodal
129+
systems such as vision-language models. Despite their remarkable success, it remains unclear how these
130+
architectures represent features internally. Here, we propose a novel approach for interpreting vision
131+
features via image reconstruction. We compare two related model families, SigLIP and SigLIP2, which differ
132+
only in their training objective, and show that encoders pre-trained on image-based tasks retain
133+
significantly more image information than those trained on non-image tasks such as contrastive learning.
134+
We further apply our method to a range of vision encoders, ranking them by the informativeness of their
135+
feature representations. Finally, we demonstrate that manipulating the feature space yields predictable
136+
changes in reconstructed images, revealing that orthogonal rotations — rather than spatial transformations
137+
— control color encoding. Our approach can be applied to any vision encoder, shedding light on the inner
138+
structure of its feature space.
139+
</div>
140+
</div>
141+
</div>
142+
</div>
143+
</section>
144+
<!-- End paper abstract -->
145+
146+
<!-- Прописать явно контрибушны: -->
147+
148+
<!-- (1) interpretability metric -->
149+
<!-- Текстовое объяснение -->
150+
<!-- Базовые результаты: siglip vs siglip2 -->
151+
<!-- Нужно четко подчеркнуть какие различия между модельками -->
152+
<!-- И как это влияет на реконструкцию -->
153+
<!-- -->
154+
155+
<!-- (2) Feature-space transformations -->
156+
<!-- Текстовое объяснение -->
157+
<!-- Визуализация фреймворка: обобщил оператор в пр-ве картинок и в пр-ве фичей -->
158+
<!-- Сделать видос с визуализацией фреймворка -->
159+
<!-- Примеры работы с RGB -->
160+
<!-- Примеры работы с отключением одного канала (ожелтением) -->
161+
<!-- -->
162+
163+
164+
<!-- Image carousel -->
165+
<section class="hero is-small">
166+
<div class="hero-body">
167+
<div class="container">
168+
<div id="results-carousel" class="carousel results-carousel">
169+
<div class="item">
170+
<!-- Your image here -->
171+
<img src="static/images/v1_vs_v2.png" alt="Comparison of SigLIP and SigLIP2 reconstructions" />
172+
<h2 class="subtitle has-text-centered">
173+
Comparison of image reconstructions from SigLIP and SigLIP2 feature spaces.
174+
</h2>
175+
</div>
176+
<div class="item">
177+
<!-- Your image here -->
178+
<img src="static/images/rb_swap.png" alt="Red-Blue channel swap visualization" />
179+
<h2 class="subtitle has-text-centered">
180+
Visualization of feature space manipulation through red-blue channel swap.
181+
</h2>
182+
</div>
183+
</div>
184+
</div>
185+
</div>
186+
</section>
187+
<!-- End image carousel -->
188+
189+
190+
191+
192+
<!--BibTex citation -->
193+
<section class="section" id="BibTeX">
194+
<div class="container is-max-desktop content">
195+
<h2 class="title">BibTeX</h2>
196+
<pre><code>@article{feature_analysis,
197+
title={Image Reconstruction as a Tool for Feature Analysis},
198+
author={Allakhverdov, Eduard and Tarasov, Dmitrii and Goncharova, Elizaveta and Kuznetsov, Andrey},
199+
journal={arXiv preprint},
200+
year={2024}
201+
}</code></pre>
202+
</div>
203+
</section>
204+
<!--End BibTex citation -->
205+
206+
207+
<footer class="footer">
208+
<div class="container">
209+
<div class="columns is-centered">
210+
<div class="column is-8">
211+
<div class="content">
212+
213+
<p>
214+
This page was built using the <a href="https://github.com/eliahuhorwitz/Academic-project-page-template"
215+
target="_blank">Academic Project Page Template</a> which was adopted from the <a
216+
href="https://nerfies.github.io" target="_blank">Nerfies</a> project page.
217+
You are free to borrow the source code of this website, we just ask that you link back to this page in the
218+
footer. <br> This website is licensed under a <a rel="license"
219+
href="http://creativecommons.org/licenses/by-sa/4.0/" target="_blank">Creative
220+
Commons Attribution-ShareAlike 4.0 International License</a>.
221+
</p>
222+
223+
</div>
224+
</div>
225+
</div>
226+
</div>
227+
</footer>
228+
229+
<!-- Statcounter tracking code -->
230+
231+
<!-- You can add a tracker to track page visits by creating an account at statcounter.com -->
232+
233+
<!-- End of Statcounter Code -->
234+
235+
</body>
236+
237+
</html>

‎project/static/css/bulma-carousel.min.css

Lines changed: 1 addition & 0 deletions
Some generated files are not rendered by default. Learn more about customizing how changed files appear on GitHub.

0 commit comments

Comments
 (0)