-
Notifications
You must be signed in to change notification settings - Fork 0
Expand file tree
/
Copy pathindex.html
More file actions
213 lines (206 loc) · 12.1 KB
/
Copy pathindex.html
File metadata and controls
213 lines (206 loc) · 12.1 KB
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
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
<!doctype html>
<html lang="en">
<head>
<meta charset="utf-8">
<meta name="viewport" content="width=device-width, initial-scale=1">
<meta name="description" content="GraphBEV++: robust multi-modal feature alignment for autonomous driving.">
<title>GraphBEV++ | Multi-Modal Feature Alignment</title>
<link rel="stylesheet" href="style.css">
</head>
<body>
<header class="site-header">
<a class="brand" href="#top" aria-label="GraphBEV++ home">
<span class="brand-mark">G<sup>++</sup></span>
<span>GraphBEV++</span>
</a>
<nav aria-label="Main navigation">
<a href="#abstract">Abstract</a>
<a href="#method">Method</a>
<a href="#results">Results</a>
<a href="#citation">Citation</a>
</nav>
<a class="nav-paper" href="paper.pdf">Paper <span>↗</span></a>
</header>
<main id="top">
<section class="hero">
<div class="hero-grid">
<div class="hero-copy">
<p class="eyebrow">Robust Multi-Modal Autonomous Driving</p>
<h1>GraphBEV<span>++</span></h1>
<p class="subtitle">Multi-Modal Feature Alignment<br>for Autonomous Driving</p>
<div class="authors">
<a href="#">Ziying Song<sup>1</sup></a>
<a href="#">Hongyu Pan<sup>2</sup></a>
<a href="#">Lin Liu<sup>1</sup></a>
<a href="#">Shaoqing Xu<sup>3</sup></a>
<a href="#">Lei Yang<sup>4</sup></a>
<a href="#">Caiyan Jia<sup>1*</sup></a>
<a href="#">Yadan Luo<sup>5</sup></a>
</div>
<p class="affiliations">
<sup>1</sup>Beijing Jiaotong University <sup>2</sup>Horizon Robotics
<sup>3</sup>University of Macau <sup>4</sup>Tsinghua University
<sup>5</sup>The University of Queensland
</p>
<div class="hero-actions">
<a class="button primary" href="paper.pdf"><span class="button-icon">PDF</span> Paper</a>
<a class="button" href="https://github.com/adept-thu/GraphBEV"><span class="button-icon"></></span> Code</a>
<a class="button" href="#citation"><span class="button-icon">“</span> BibTeX</a>
</div>
</div>
<aside class="hero-visual" aria-label="GraphBEV++ conceptual summary">
<div class="visual-label">A unified alignment framework</div>
<img src="resources/overview.png" alt="GraphBEV++ overview from misalignment to end-to-end driving">
</aside>
</div>
<div class="metric-strip">
<div><strong>+8.5</strong><span>mAP vs. BEVFusion<br>under noise</span></div>
<div><strong>69.1</strong><span>mAP on<br>nuScenes-C</span></div>
<div><strong>51.1</strong><span>AMOTA for<br>end-to-end tracking</span></div>
<div><strong>88.7</strong><span>PDMS on<br>NAVSIM</span></div>
</div>
</section>
<section id="abstract" class="section abstract-section">
<div class="section-index">01</div>
<div>
<p class="section-kicker">Motivation</p>
<h2>Alignment is a prerequisite<br>for reliable fusion.</h2>
</div>
<div class="abstract-copy">
<p class="lead">Feature misalignment in BEV perception is a critical yet often overlooked challenge, particularly under calibration uncertainty between LiDAR and camera sensors.</p>
<p>GraphBEV++ systematically mitigates projection-induced misalignment with two complementary modules. <strong>LocalAlign-v2</strong> corrects local correspondence errors through neighborhood-aware graph matching, while <strong>GlobalAlign-v2</strong> resolves representation-level inconsistencies through deformable offset learning or diffusion-based denoising.</p>
<p>The framework supports both LSS-based and query-based BEV representations, generalizes from detection to occupancy prediction, and improves perception, prediction, and planning in end-to-end autonomous driving.</p>
<div class="keywords">
<span>Autonomous Driving</span><span>Multi-Modal Fusion</span><span>Feature Alignment</span><span>Bird's-Eye View</span>
</div>
</div>
</section>
<section id="method" class="section method-section">
<div class="section-heading">
<div><span class="section-index">02</span><p class="section-kicker">Method</p></div>
<h2>Local correspondence.<br>Global consistency.</h2>
<p>GraphBEV++ treats misalignment as a hierarchical error propagation process, correcting it at both BEV construction and fusion stages.</p>
</div>
<figure class="paper-figure wide-figure">
<img src="resources/framework.png" alt="GraphBEV++ end-to-end autonomous driving framework">
<figcaption><span>Figure 2.</span> GraphBEV++ within a multi-modal end-to-end autonomous driving framework.</figcaption>
</figure>
<div class="method-grid">
<article class="method-card local">
<div class="card-number">A</div>
<p class="card-tag">BEV Construction</p>
<h3>LocalAlign-v2</h3>
<p>Builds neighborhood-aware representations to compensate for inaccurate LiDAR-to-camera projections and reference-point deviations.</p>
<ul>
<li><strong>LSS variant</strong> encodes projected and neighboring depth.</li>
<li><strong>Query variant</strong> refines BEV queries using adjacent queries.</li>
<li><strong>Adaptive KNN</strong> allocates context by object scale and depth.</li>
</ul>
</article>
<article class="method-card global">
<div class="card-number">B</div>
<p class="card-tag">BEV Fusion</p>
<h3>GlobalAlign-v2</h3>
<p>Aligns heterogeneous BEV representations after local errors accumulate into spatial shifts and semantic inconsistencies.</p>
<ul>
<li><strong>Deformable variant</strong> learns explicit spatial offsets.</li>
<li><strong>Diffusion variant</strong> progressively denoises implicit features.</li>
<li><strong>Four-step refinement</strong> balances robustness and efficiency.</li>
</ul>
</article>
</div>
<div class="figure-pair">
<figure class="paper-figure"><img src="resources/local-align.png" alt="LocalAlign-v2 LSS and Query pipelines"><figcaption>LocalAlign-v2 for LSS and Query representations.</figcaption></figure>
<figure class="paper-figure"><img src="resources/global-align.png" alt="GlobalAlign-v2 Deformable and Diffusion pipelines"><figcaption>GlobalAlign-v2 with deformable and diffusion alignment.</figcaption></figure>
</div>
</section>
<section id="results" class="section results-section">
<div class="section-heading compact">
<div><span class="section-index">03</span><p class="section-kicker">Experiments</p></div>
<h2>Robust across tasks,<br>datasets, and noise.</h2>
<p>Evaluation spans 3D detection, BEV segmentation, semantic occupancy, and end-to-end driving under clean and misaligned settings.</p>
</div>
<div class="tabs" role="tablist" aria-label="Result tables">
<button class="tab active" data-table="detection" role="tab">Detection robustness</button>
<button class="tab" data-table="e2e" role="tab">End-to-end driving</button>
<button class="tab" data-table="generalization" role="tab">Generalization</button>
</div>
<div class="table-shell active" id="detection">
<table>
<thead><tr><th>Method</th><th>Clean mAP ↑</th><th>Noisy mAP ↑</th><th>Clean NDS ↑</th><th>Noisy NDS ↑</th><th>Relative mAP drop ↓</th></tr></thead>
<tbody>
<tr><td>SparseFusion</td><td>70.4</td><td>64.7</td><td>72.8</td><td>67.1</td><td>8.1%</td></tr>
<tr><td>BEVFusion-MIT</td><td>68.5</td><td>60.8</td><td>71.4</td><td>65.7</td><td>11.2%</td></tr>
<tr><td>BEVFormer-M</td><td>70.9</td><td>63.2</td><td>73.0</td><td>66.3</td><td>10.8%</td></tr>
<tr class="highlight"><td>GraphBEV++ (LSS)</td><td>70.7</td><td>69.3</td><td>73.2</td><td>72.3</td><td><strong>2.0%</strong></td></tr>
<tr class="highlight"><td>GraphBEV++ (Query)</td><td><strong>71.4</strong></td><td>69.1</td><td><strong>73.4</strong></td><td>71.2</td><td>3.2%</td></tr>
</tbody>
</table>
</div>
<div class="table-shell" id="e2e">
<table>
<thead><tr><th>Method</th><th>AMOTA ↑</th><th>minADE ↓</th><th>minFDE ↓</th><th>MR ↓</th><th>EPA ↑</th><th>Avg. collision ↓</th></tr></thead>
<tbody>
<tr><td>UniAD</td><td>35.9</td><td>0.71</td><td>1.02</td><td>15.1</td><td>45.6</td><td>0.31</td></tr>
<tr><td>FusionAD</td><td>50.1</td><td>0.39</td><td>0.62</td><td>8.6</td><td>62.6</td><td>0.12</td></tr>
<tr class="highlight"><td>GraphBEV++ (LSS)</td><td>49.8</td><td>0.40</td><td>0.59</td><td>8.5</td><td><strong>64.7</strong></td><td>0.21</td></tr>
<tr class="highlight"><td>GraphBEV++ (Query)</td><td><strong>51.1</strong></td><td><strong>0.38</strong></td><td><strong>0.52</strong></td><td><strong>7.7</strong></td><td>64.5</td><td>0.13</td></tr>
</tbody>
</table>
</div>
<div class="table-shell" id="generalization">
<table>
<thead><tr><th>Benchmark / task</th><th>Baseline</th><th>GraphBEV++</th><th>Key metric</th><th>Gain</th></tr></thead>
<tbody>
<tr><td>Waymo-C L2</td><td>32.44 / 29.95</td><td>36.48 / 33.08</td><td>mAP / mAPH</td><td>+4.04 / +3.13</td></tr>
<tr><td>Argoverse2</td><td>43.1</td><td><strong>46.7</strong></td><td>mAP</td><td>+3.6</td></tr>
<tr><td>3D occupancy, noisy</td><td>27.86 / 17.63</td><td><strong>29.41 / 19.37</strong></td><td>IoU / mIoU</td><td>+1.55 / +1.74</td></tr>
<tr><td>NAVSIM</td><td>88.3</td><td><strong>88.7</strong></td><td>PDMS</td><td>+0.4</td></tr>
</tbody>
</table>
</div>
<figure class="paper-figure robustness-figure">
<img src="resources/robustness.png" alt="Performance with respect to misalignment severity">
<figcaption><span>Robustness analysis.</span> GraphBEV++ degrades substantially more gracefully as misalignment severity increases.</figcaption>
</figure>
</section>
<section class="section takeaways-section">
<div class="section-heading compact">
<div><span class="section-index">04</span><p class="section-kicker">Key findings</p></div>
<h2>What the experiments establish.</h2>
</div>
<div class="takeaways">
<article><span>01</span><h3>Cross-paradigm</h3><p>A single alignment principle extends to dense LSS-based and sparse query-based BEV representations.</p></article>
<article><span>02</span><h3>Noise-resilient</h3><p>Under nuScenes-C, the LSS variant retains 98% of its clean-setting mAP.</p></article>
<article><span>03</span><h3>Task-general</h3><p>Better geometric correspondence improves detection, tracking, forecasting, occupancy, and planning.</p></article>
<article><span>04</span><h3>Efficient</h3><p>GraphBEV++ (LSS) runs at 7.1 FPS with only marginal alignment overhead on an A100 GPU.</p></article>
</div>
</section>
<section id="citation" class="section citation-section">
<div>
<span class="section-index">05</span>
<p class="section-kicker">Citation</p>
<h2>Cite GraphBEV++</h2>
<p>If this work supports your research, please cite the paper.</p>
</div>
<div class="bibtex-wrap">
<button id="copyBibtex">Copy BibTeX</button>
<pre id="bibtex">@misc{song2026graphbevplusplus,
title = {GraphBEV++: Multi-Modal Feature Alignment
for Autonomous Driving},
author = {Song, Ziying and Pan, Hongyu and Liu, Lin and
Xu, Shaoqing and Yang, Lei and Jia, Caiyan and Luo, Yadan},
note = {Manuscript},
year = {2026}
}</pre>
</div>
</section>
</main>
<footer>
<a class="brand" href="#top"><span class="brand-mark">G<sup>++</sup></span><span>GraphBEV++</span></a>
<p>Multi-Modal Feature Alignment for Autonomous Driving</p>
<p class="footer-note">Academic project page · 2026</p>
</footer>
<script src="script.js"></script>
</body>
</html>