Skip to content

Commit 671c5bb

Browse files
egorchistovegorchistov
committed
Add page
1 parent ac59957 commit 671c5bb

29 files changed

Lines changed: 1417 additions & 0 deletions

‎docs/index.html‎

Lines changed: 334 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,334 @@
1+
<!DOCTYPE html>
2+
<html lang="en">
3+
4+
<head>
5+
<meta charset="utf-8">
6+
<meta name="description"
7+
content="MEMFOF is a memory-efficient optical flow method for Full HD video that combines high accuracy with low VRAM usage.">
8+
<meta name="keywords" content="optical flow, memory-efficient, high-resolution">
9+
<meta name="viewport" content="width=device-width, initial-scale=1">
10+
<title>MEMFOF: High-Resolution Training for Memory-Efficient Multi-Frame Optical Flow Estimation</title>
11+
12+
<link href="https://fonts.googleapis.com/css?family=Google+Sans|Noto+Sans|Castoro" rel="stylesheet">
13+
14+
<link rel="stylesheet" href="static/css/bulma.min.css">
15+
<link rel="stylesheet" href="static/css/index.css">
16+
<link rel="icon" href="static/images/favicon.svg">
17+
18+
<link rel="stylesheet" href="https://unpkg.com/img-comparison-slider/dist/styles.css">
19+
20+
<script defer src="static/js/comparison-slider.js"></script>
21+
<script defer src="static/js/fontawesome.all.min.js"></script>
22+
<script defer src="static/js/kitti-chart.js"></script>
23+
<script defer src="static/js/sintel-chart.js"></script>
24+
<script defer src="static/js/spring-chart.js"></script>
25+
26+
<script defer src="https://unpkg.com/img-comparison-slider/dist/index.js"></script>
27+
<script defer src="https://cdn.jsdelivr.net/npm/chart.js"></script>
28+
<script defer src="https://cdn.jsdelivr.net/npm/chartjs-plugin-datalabels"></script>
29+
</head>
30+
31+
<body>
32+
33+
<section class="hero" id="authors">
34+
<div class="hero-body">
35+
<div class="container is-max-desktop">
36+
<div class="columns is-centered">
37+
<div class="column has-text-centered">
38+
<h1 class="title is-1 publication-title">High-Resolution Training for Memory-Efficient Multi-Frame Optical
39+
Flow</h1>
40+
<div class="is-size-5 publication-authors">
41+
<span class="author-block">
42+
<a href="https://github.com/VladBargin">Vladislav Bargatin</a><sup>1</sup>
43+
</span>
44+
<span class="author-block">
45+
<a href="http://github.com/egorchistov">Egor Chistov</a><sup>1</sup>
46+
</span>
47+
<span class="author-block">
48+
<a href="https://github.com/AlexanderYakovenko1">Alexander
49+
Yakovenko</a><sup>1,2</sup>
50+
</span>
51+
<span class="author-block">
52+
<a href="https://linkedin.com/in/dmitriyvatolin">Dmitriy Vatolin</a><sup>1,2</sup>
53+
</span>
54+
</div>
55+
56+
<div class="is-size-5 publication-authors">
57+
<span class="author-block"><sup>1</sup>Lomonosov Moscow State University</span>
58+
<span class="author-block"><sup>2</sup>MSU Institute for Artificial Intelligence</span>
59+
</div>
60+
61+
<div class="column has-text-centered">
62+
<div class="publication-links">
63+
<span class="link-block">
64+
<a href="https://arxiv.org/abs/2506.23151" class="external-link button is-normal is-rounded is-dark">
65+
<span class="icon">
66+
<i class="fas fa-file-pdf"></i>
67+
</span>
68+
<span>Paper</span>
69+
</a>
70+
</span>
71+
<span class="link-block">
72+
<a href="https://github.com/msu-video-group/memfof" class="external-link button is-normal is-rounded is-dark">
73+
<span class="icon">
74+
<i class="fab fa-github"></i>
75+
</span>
76+
<span>Code</span>
77+
</a>
78+
</span>
79+
<span class="link-block">
80+
<a href="https://huggingface.co/spaces/egorchistov/MEMFOF" class="external-link button is-normal is-rounded is-dark">
81+
<span class="icon">
82+
<i class="far fa-images"></i>
83+
</span>
84+
<span>Demo</span>
85+
</a>
86+
</span>
87+
<span class="link-block">
88+
<a href="https://colab.research.google.com/github/msu-video-group/memfof/blob/dev/demo.ipynb" class="external-link button is-normal is-rounded is-dark">
89+
<span class="icon">
90+
<i class="far fa-images"></i>
91+
</span>
92+
<span>Colab</span>
93+
</a>
94+
</span>
95+
</div>
96+
97+
</div>
98+
</div>
99+
</div>
100+
</div>
101+
</div>
102+
</section>
103+
104+
<section class="section" id="spring-videos">
105+
<div class="container is-max-desktop has-text-centered">
106+
<div class="columns is-centered">
107+
<div class="column is-two-thirds">
108+
<h2 class="title is-3">Native-Resolution Results on Spring</h2>
109+
<p>MEMFOF demonstrates that multi-frame processing <b>enhances temporal coherence</b>,
110+
and native resolution input helps <b>retain details</b></p>
111+
</div>
112+
</div>
113+
<img-comparison-slider>
114+
<video slot="first" src="static/videos/spring_flow.mp4" muted loop autoplay></video>
115+
<video slot="second" src="static/videos/spring_input.mp4" muted loop autoplay></video>
116+
</img-comparison-slider>
117+
<p>
118+
Controls: <b>Click</b> or press <b>spacebar</b> to play/pause; <b>drag slider</b> to compare results; use
119+
<b>←</b>/<b>→</b> <b>arrows</b> to step through frames.
120+
</p>
121+
</div>
122+
</section>
123+
124+
<section class="section" id="highlights">
125+
<div class="container is-max-desktop has-text-centered">
126+
<div class="columns is-centered">
127+
<div class="column is-two-thirds">
128+
<h2 class="title is-3">Highlights</h2>
129+
<div class="content has-text-justified">
130+
<ul>
131+
<li>MEMFOF delivers high-accuracy optical flow estimation for Full HD video while significantly reducing
132+
GPU
133+
memory usage — requiring <b>just 2.09 GB</b> — enabling <strong>native 1080p processing</strong> without
134+
cropping or downsampling</li>
135+
<li>By combining multi-frame estimation, scalable correlation volumes, and resolution-aware training, it
136+
achieves <b>state-of-the-art results across multiple benchmarks</b> with lower resource demands</li>
137+
<li>The method <b>ranks first on Spring and Sintel (Clean) benchmarks</b> and shows
138+
strong performance on <b>KITTI-2015</b>, combining superior accuracy with efficiency
139+
</li>
140+
</ul>
141+
</div>
142+
</div>
143+
</div>
144+
</div>
145+
</section>
146+
147+
<section class="section" id="histograms">
148+
<div class="container is-max-desktop has-text-centered">
149+
<div class="columns is-centered">
150+
<div class="column is-two-thirds">
151+
<h2 class="title is-3">High-Resolution Data</h2>
152+
<p>Using 2D histograms, we <b>analyze and compare motion patterns</b> across optical flow datasets to uncover
153+
the full range and distribution of movements. We <b>identify gaps in existing data</b> and <b>enhance
154+
training set through upsampling</b> to address them</p>
155+
</div>
156+
</div>
157+
<div class="is-flex is-flex-wrap-wrap is-justify-content-center">
158+
<!-- 8 ячеек -->
159+
<div class="item">
160+
<image src="static/images/hists/tartan.png"></image>
161+
<p class="has-text-weight-semibold mt-2">TartanAir</p>
162+
</div>
163+
<div class="item">
164+
<image src="static/images/hists/things.png"></image>
165+
<p class="has-text-weight-semibold mt-2">FlyingThings</p>
166+
</div>
167+
<div class="item">
168+
<image src="static/images/hists/kitti.png"></image>
169+
<p class="has-text-weight-semibold mt-2">KITTI-2015</p>
170+
</div>
171+
<div class="item">
172+
<image src="static/images/hists/hd1k.png"></image>
173+
<p class="has-text-weight-semibold mt-2">HD1K</p>
174+
</div>
175+
<div class="item">
176+
<image src="static/images/hists/sintel.png"></image>
177+
<p class="has-text-weight-semibold mt-2">Sintel</p>
178+
</div>
179+
<div class="item">
180+
<image src="static/images/hists/spring.png"></image>
181+
<p class="has-text-weight-semibold mt-2">Spring</p>
182+
</div>
183+
<div class="item">
184+
<image src="static/images/hists/combined.png"></image>
185+
<p class="has-text-weight-semibold mt-2">Combined</p>
186+
</div>
187+
<div class="item">
188+
<image src="static/images/hists/combined2x.png"></image>
189+
<p class="has-text-weight-semibold mt-2">Combined at 2x resolution</p>
190+
</div>
191+
</div>
192+
<p>Color intensity indicates the number of motion vectors per bin, with borders marking each dataset’s maximum
193+
motion range. Large motions in the Spring dataset, missing from other training sets, are captured after 2x
194+
upsampling the combined data</p>
195+
</div>
196+
</section>
197+
198+
<section class="section" id="memory-efficiency">
199+
<div class="container is-max-desktop has-text-centered">
200+
<div class="columns is-centered">
201+
<div class="column is-two-thirds">
202+
<h2 class="title is-3">Memory Efficiency</h2>
203+
<p>By reducing the resolution of correlation volumes, we free up enough memory to implement a multi-frame
204+
method that improves temporal stability. Our method requires <b>just 2.09 GB of memory for inference,</b>
205+
enabling <b>native Full HD training</b></p>
206+
</div>
207+
</div>
208+
209+
<div class="grid">
210+
<div class="item">
211+
<object type="image/svg+xml" data="static/images/one-corr.svg"></object>
212+
<p class="m-4 has-text-weight-semibold">SEA-RAFT</p>
213+
</div>
214+
<div class="item">
215+
<object type="image/svg+xml" data="static/images/two-corrs.svg"></object>
216+
<p class="m-4 has-text-weight-semibold">MEMFOF</p>
217+
</div>
218+
</div>
219+
220+
<p>
221+
Reducing correlation volume resolution lowers memory use but can degrade quality. Our three-frame approach
222+
compensates for this, restoring accuracy while keeping efficiency and enabling native Full HD processing
223+
</p>
224+
</div>
225+
</section>
226+
227+
<section class="section" id="error-vs-memory">
228+
<div class="container is-max-desktop has-text-centered">
229+
<div class="columns is-centered">
230+
<div class="column is-two-thirds">
231+
<h2 class="title is-3">1px Error vs. Memory Usage on Spring</h2>
232+
<p>MEMFOF demonstrates <b>superior memory efficiency</b> and the <b>lowest error</b> among all methods. Speed
233+
and peak memory usage were measured on a Nvidia RTX 3090</p>
234+
</div>
235+
</div>
236+
<div>
237+
<canvas id="spring-chart"></canvas>
238+
</div>
239+
<p>Results are sourced from <a href="https://spring-benchmark.org/opticalflow">official leaderboard</a> of the
240+
Spring benchmark.<br>w/o ft stands for methods that were not finetuned on Spring dataset</p>
241+
</div>
242+
</section>
243+
244+
<section class="section" id="sintel-videos">
245+
<div class="container is-max-desktop has-text-centered">
246+
<div class="columns is-centered">
247+
<div class="column is-two-thirds">
248+
<h2 class="title is-3">Performance on Sintel Benchmark</h2>
249+
<p>MEMFOF produces <b>consistent motion</b> across large deformations and occlusions in challenging cinematic scenes</p>
250+
</div>
251+
</div>
252+
<img-comparison-slider>
253+
<video slot="first" src="static/videos/sintel_flow.mp4" muted loop autoplay></video>
254+
<video slot="second" src="static/videos/sintel_input.mp4" muted loop autoplay></video>
255+
</img-comparison-slider>
256+
<p>
257+
Controls: <b>Click</b> or press <b>spacebar</b> to play/pause; <b>drag slider</b> to compare results; use
258+
<b>←</b>/<b>→</b> <b>arrows</b> to step through frames.
259+
</p>
260+
</div>
261+
</section>
262+
263+
<section class="section" id="sintel-results">
264+
<div class="container is-max-desktop has-text-centered">
265+
<div class="columns is-centered">
266+
<div class="column is-two-thirds">
267+
<h2 class="title is-3">EndPoint Error on Sintel</h2>
268+
<p>MEMFOF achieves competitive performance on both Sintel splits, sharing <b>first place</b> with the five-frame version of VideoFlow <b>on the clean pass</b> and outperforming SEA-RAFT (L) by 32% on the final pass</p>
269+
</div>
270+
</div>
271+
<div>
272+
<canvas id="sintel-chart"></canvas>
273+
</div>
274+
<p>Results are sourced from <a href="http://sintel.is.tue.mpg.de/quant?metric_id=0&selected_pass=1">official leaderboard</a> of the Sintel benchmark.<br>Methods are sorted by performance on Sintel Clean split</p>
275+
</div>
276+
</section>
277+
278+
<section class="section" id="kitti-videos">
279+
<div class="container is-max-desktop has-text-centered">
280+
<div class="columns is-centered">
281+
<div class="column is-two-thirds">
282+
<h2 class="title is-3">Performance in Real-World KITTI Scenes</h2>
283+
<p>MEMFOF excels in real-world driving scenes, showing <b>high accuracy and stability</b> across challenging motions and lighting</b></p>
284+
</div>
285+
</div>
286+
<img-comparison-slider>
287+
<video slot="first" src="static/videos/kitti_flow.mp4" muted loop autoplay></video>
288+
<video slot="second" src="static/videos/kitti_input.mp4" muted loop autoplay></video>
289+
</img-comparison-slider>
290+
<p>
291+
Controls: <b>Click</b> or press <b>spacebar</b> to play/pause; <b>drag slider</b> to compare results; use
292+
<b>←</b>/<b>→</b> <b>arrows</b> to step through frames.
293+
</p>
294+
</div>
295+
</section>
296+
297+
<section class="section" id="kitti-results">
298+
<div class="container is-max-desktop has-text-centered">
299+
<div class="columns is-centered">
300+
<div class="column is-two-thirds">
301+
<h2 class="title is-3">Fl-all on KITTI-2015</h2>
302+
<p>MEMFOF achieves <b>state-of-the-art performance</b> among all non-scene flow methods, outperforming both SEA-RAFT and VideoFlow</p>
303+
</div>
304+
</div>
305+
<div>
306+
<canvas id="kitti-chart"></canvas>
307+
</div>
308+
<p>Results are sourced from <a href="https://www.cvlibs.net/datasets/kitti/eval_scene_flow.php?benchmark=flow">official leaderboard</a> of the KITTI benchmark. <br>Only non-scene flow methods are shown</p>
309+
</div>
310+
</section>
311+
312+
<section class="section" id="bibtex">
313+
<div class="container is-max-desktop content">
314+
<h2 class="title">BibTeX</h2>
315+
<pre><code>@article{bargatin2025memfof,
316+
title = {MEMFOF: High-Resolution Training for Memory-Efficient Multi-Frame Optical Flow Estimation},
317+
author = {Bargatin, Vladislav and Chistov, Egor and Yakovenko, Alexander and Vatolin, Dmitriy},
318+
journal = {arXiv preprint arXiv:2506.23151},
319+
year = {2025}
320+
}</code></pre>
321+
</div>
322+
</section>
323+
324+
<section class="section" id="credit">
325+
<div class="content has-text-centered">
326+
<p>
327+
The project page is based on <a href="https://github.com/nerfies/nerfies.github.io">this template</a>,
328+
licensed under a <a href="http://creativecommons.org/licenses/by-sa/4.0/">CC BY-SA 4.0</a> license
329+
</p>
330+
</div>
331+
</section>
332+
</body>
333+
334+
</html>

‎docs/static/css/bulma.css.map.txt‎

Lines changed: 1 addition & 0 deletions
Large diffs are not rendered by default.

‎docs/static/css/bulma.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.

‎docs/static/css/fontawesome.all.min.css‎

Lines changed: 5 additions & 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)