-
Notifications
You must be signed in to change notification settings - Fork 0
Expand file tree
/
Copy pathindex.html
More file actions
1101 lines (996 loc) · 48.7 KB
/
Copy pathindex.html
File metadata and controls
1101 lines (996 loc) · 48.7 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
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
<!DOCTYPE html>
<html lang="en">
<head>
<meta charset="UTF-8">
<meta name="viewport" content="width=device-width, initial-scale=1.0">
<title>How Well Do Vision-Language Models Understand Sequential Driving Scenes?</title>
<link rel="preconnect" href="https://fonts.googleapis.com">
<link rel="preconnect" href="https://fonts.gstatic.com" crossorigin>
<link href="https://fonts.googleapis.com/css2?family=Noto+Sans:wght@400;500;600;700&display=swap" rel="stylesheet">
<link rel="stylesheet" href="https://cdnjs.cloudflare.com/ajax/libs/font-awesome/6.5.1/css/all.min.css">
<script src="https://cdn.plot.ly/plotly-basic-2.35.2.min.js"></script>
<style>
* { margin: 0; padding: 0; box-sizing: border-box; }
html { scroll-behavior: smooth; }
body {
font-family: 'Noto Sans', sans-serif;
color: #333;
background: #fff;
line-height: 1.65;
font-size: 1rem;
}
/* ===== Hero ===== */
.hero {
background: #fff;
padding: 4rem 2rem 2.5rem;
text-align: center;
max-width: 960px;
margin: 0 auto;
}
.hero h1 {
font-size: 2.8rem;
font-weight: 700;
color: #222;
letter-spacing: -0.5px;
margin-bottom: 0.3rem;
}
.hero .acronym {
font-size: 1.05rem;
color: #555;
margin-bottom: 1.2rem;
}
.hero .authors {
font-size: 0.95rem;
color: #555;
margin-bottom: 0.3rem;
}
.hero .affiliations {
font-size: 0.85rem;
color: #888;
margin-bottom: 1.5rem;
}
.hero-links {
display: flex;
gap: 0.6rem;
justify-content: center;
flex-wrap: wrap;
}
/* Pill buttons */
.btn-pill {
display: inline-flex;
align-items: center;
gap: 0.4rem;
padding: 0.5rem 1.3rem;
border-radius: 20px;
text-decoration: none;
font-size: 0.88rem;
font-weight: 500;
transition: background 0.2s, box-shadow 0.2s;
border: none;
cursor: pointer;
}
.btn-dark {
background: #363636;
color: #fff;
}
.btn-dark:hover {
background: #222;
box-shadow: 0 4px 14px rgba(0,0,0,0.15);
}
.btn-disabled {
background: #d5d5d5;
color: #888;
cursor: default;
}
/* ===== Container ===== */
.container {
max-width: 960px;
margin: 0 auto;
padding: 0 2rem;
}
/* ===== Sections ===== */
section {
padding: 2.5rem 0;
}
section + section {
border-top: 1px solid #eee;
}
/* ===== Typography ===== */
h2 {
font-size: 1.8rem;
font-weight: 600;
color: #222;
margin-bottom: 1rem;
}
h3 {
font-size: 1.3rem;
font-weight: 600;
color: #333;
margin: 1.5rem 0 0.6rem;
}
h4 {
font-size: 1.05rem;
font-weight: 600;
color: #444;
margin: 1rem 0 0.4rem;
}
p { margin-bottom: 0.75rem; }
ul, ol { margin: 0.5rem 0 0.8rem 1.5rem; }
li { margin-bottom: 0.3rem; }
a { color: #4a6fa5; }
/* ===== Stat cards ===== */
.card-grid {
display: grid;
grid-template-columns: repeat(auto-fit, minmax(200px, 1fr));
gap: 1rem;
margin: 1rem 0;
}
.stat-card {
background: #fafafa;
border: 1px solid #eee;
border-radius: 8px;
padding: 1.2rem;
text-align: center;
}
.stat-card .number {
font-size: 2rem;
font-weight: 700;
color: #4a6fa5;
}
.stat-card .label {
font-size: 0.85rem;
color: #666;
margin-top: 0.2rem;
}
/* ===== Figures ===== */
.figure {
margin: 1.5rem 0;
text-align: center;
}
.figure img {
max-width: 100%;
border-radius: 4px;
}
.figure figcaption, .figure .caption {
font-size: 0.85rem;
color: #666;
margin-top: 0.5rem;
font-style: italic;
max-width: 720px;
margin-left: auto;
margin-right: auto;
}
/* ===== Code blocks ===== */
pre {
background: #f5f5f5;
border: 1px solid #e8e8e8;
border-radius: 6px;
padding: 1rem;
overflow-x: auto;
font-size: 0.8rem;
line-height: 1.5;
margin: 0.8rem 0;
}
code {
font-family: 'SFMono-Regular', 'Consolas', 'Liberation Mono', 'Menlo', monospace;
font-size: 0.8rem;
}
p code, li code {
background: #f5f5f5;
padding: 0.15rem 0.4rem;
border-radius: 3px;
font-size: 0.82em;
}
.code-label {
font-size: 0.78rem;
font-weight: 600;
color: #888;
text-transform: uppercase;
letter-spacing: 0.5px;
margin-bottom: 0.3rem;
}
/* ===== Collapsible ===== */
details {
margin: 0.8rem 0;
border: 1px solid #e8e8e8;
border-radius: 8px;
overflow: hidden;
}
summary {
padding: 0.8rem 1rem;
background: #fafafa;
cursor: pointer;
font-weight: 600;
font-size: 0.95rem;
user-select: none;
list-style: none;
display: flex;
align-items: center;
gap: 0.5rem;
transition: background 0.15s;
}
summary:hover { background: #f2f2f2; }
summary::before {
content: '\25B6';
font-size: 0.65rem;
color: #888;
transition: transform 0.2s;
}
details[open] > summary::before {
transform: rotate(90deg);
}
summary::-webkit-details-marker { display: none; }
details > .details-content {
padding: 1rem;
border-top: 1px solid #e8e8e8;
}
/* ===== Tables ===== */
table {
width: 100%;
border-collapse: collapse;
margin: 1rem 0;
font-size: 0.82rem;
}
th, td {
padding: 0.45rem 0.6rem;
border: 1px solid #e8e8e8;
text-align: center;
}
th {
background: #f5f5f5;
color: #333;
font-weight: 600;
}
tr:nth-child(even) { background: #fafafa; }
.table-wrapper {
overflow-x: auto;
}
/* ===== Supplementary badge ===== */
.supp-badge {
display: inline-block;
background: #f0ad4e;
color: white;
padding: 0.15rem 0.5rem;
border-radius: 4px;
font-size: 0.72rem;
font-weight: 700;
vertical-align: middle;
margin-right: 0.3rem;
}
/* ===== Difficulty badges ===== */
.badge {
display: inline-block;
padding: 0.15rem 0.5rem;
border-radius: 4px;
font-size: 0.75rem;
font-weight: 600;
}
.badge-easy { background: #d4edda; color: #155724; }
.badge-moderate { background: #fff3cd; color: #856404; }
.badge-hard { background: #f8d7da; color: #721c24; }
/* ===== Citation ===== */
.citation-block {
background: #f5f5f5;
border: 1px solid #e8e8e8;
border-radius: 6px;
padding: 1rem;
font-family: monospace;
font-size: 0.8rem;
white-space: pre-wrap;
position: relative;
}
.copy-btn {
position: absolute;
top: 0.5rem;
right: 0.5rem;
padding: 0.3rem 0.6rem;
border: 1px solid #ddd;
border-radius: 4px;
background: white;
cursor: pointer;
font-size: 0.75rem;
transition: background 0.15s;
}
.copy-btn:hover { background: #f5f5f5; }
/* ===== Footer ===== */
footer {
text-align: center;
padding: 2rem;
font-size: 0.85rem;
color: #999;
border-top: 1px solid #eee;
margin-top: 1rem;
}
/* ===== Responsive ===== */
@media (max-width: 768px) {
.hero { padding: 2.5rem 1.5rem 2rem; }
.hero h1 { font-size: 2rem; }
.hero-links { flex-direction: column; align-items: center; }
.card-grid { grid-template-columns: 1fr 1fr; }
.container { padding: 0 1.2rem; }
}
@media (max-width: 480px) {
.card-grid { grid-template-columns: 1fr; }
}
</style>
</head>
<body>
<!-- ============================================================ -->
<!-- HERO -->
<!-- ============================================================ -->
<div class="hero">
<h1>How Well Do Vision-Language Models Understand Sequential Driving Scenes? A Sensitivity Study</h1>
<div class="acronym">VENUSS: VLM Evaluation oN Understanding Sequential Scenes</div>
<div class="authors">Roberto Brusnicki, Mattia Piccinini, Johannes Betz</div>
<div class="affiliations">Technical University of Munich</div>
<div class="hero-links">
<a href="https://arxiv.org/abs/2604.06750" class="btn-pill btn-dark"><i class="fa-solid fa-file-lines"></i> Paper</a>
<a href="https://github.com/TUM-AVS/VENUSS" class="btn-pill btn-dark"><i class="fa-brands fa-github"></i> Code</a>
<a href="demo/" class="btn-pill btn-dark"><i class="fa-solid fa-gamepad"></i> Try the Demo</a>
</div>
</div>
<div class="container">
<!-- ============================================================ -->
<!-- ABSTRACT -->
<!-- ============================================================ -->
<section id="abstract">
<h2>Abstract</h2>
<p>Vision-Language Models (VLMs) are increasingly proposed for autonomous driving tasks, yet their performance on sequential driving scenes remains poorly characterized, particularly regarding how input configurations affect their capabilities.</p>
<p>We introduce <strong>VENUSS</strong>, a framework for systematic sensitivity analysis of VLM performance on sequential driving scenes, establishing baselines for future research. Building upon existing datasets, VENUSS extracts temporal sequences from driving videos and generates structured evaluations across custom categories.</p>
<p>By comparing 25+ existing VLMs across 2,600+ scenarios, we reveal how even top models achieve only <strong>57% accuracy</strong>, not matching human performance in similar constraints (65%) and exposing significant capability gaps. Our analysis shows that VLMs excel with static object detection but struggle with understanding the vehicle dynamics and temporal relations.</p>
<p>VENUSS offers the first systematic sensitivity analysis of VLMs focused on how input image configurations – resolution, frame count, temporal intervals, spatial layouts, and presentation modes – affect performance on sequential driving scenes.</p>
</section>
<!-- ============================================================ -->
<!-- FRAMEWORK OVERVIEW -->
<!-- ============================================================ -->
<section id="framework">
<h2>Framework Overview</h2>
<div class="figure">
<img src="assets/venuss_framework_overview.png" alt="VENUSS Framework Overview">
<div class="caption">VENUSS framework overview. Starting from driving datasets, VENUSS generates structured evaluation data with controlled variations in image count, timing, resolution, layout, and presentation mode. It evaluates both VLMs and humans on identical tasks across custom categories, identifies optimal input configurations per model, and establishes performance baselines.</div>
</div>
<h3>Four Modules</h3>
<div class="card-grid">
<div class="stat-card">
<div class="number" style="font-size:1.3rem;">M1</div>
<div class="label"><strong>Dataset Generation</strong><br>Extract temporal sequences from driving videos with controlled variations</div>
</div>
<div class="stat-card">
<div class="number" style="font-size:1.3rem;">M2</div>
<div class="label"><strong>Annotation System</strong><br>Transform textual descriptions into structured, fine-grained labels</div>
</div>
<div class="stat-card">
<div class="number" style="font-size:1.3rem;">M3</div>
<div class="label"><strong>Human Baselines</strong><br>Web application for human evaluation with collage, GIF, and video modes</div>
</div>
<div class="stat-card">
<div class="number" style="font-size:1.3rem;">M4</div>
<div class="label"><strong>Prompt Design</strong><br>Standardized prompts for fair VLM comparison across configurations</div>
</div>
</div>
<h3>Dataset-Agnostic Design</h3>
<p>VENUSS is designed to be dataset-agnostic. Categories are automatically extracted from each dataset's textual descriptions. We release VENUSS with configurations for four datasets:</p>
<ul>
<li><strong>CoVLA</strong>: 7 <em>behavioral</em> categories (motion, direction, speed, following, acceleration, traffic lights, curvature) – extracted from natural language captions.</li>
<li><strong>Honda Scenes</strong>: 16 <em>environmental</em> categories (road type, weather, surface, ambient lighting, infrastructure) – derived from categorical CSV annotations.</li>
<li><strong>Waymo Open Dataset</strong>: Multi-sensor driving data with LiDAR and camera annotations – <em>coming soon</em>.</li>
<li><strong>NuScenes</strong>: Multi-sensor urban driving data with 3D object annotations – <em>coming soon</em>.</li>
</ul>
<p>Adding a new dataset requires modifying only 3 files: dataset configuration, annotation parser, and evaluation questions.</p>
</section>
<!-- ============================================================ -->
<!-- ANNOTATION SYSTEM -->
<!-- ============================================================ -->
<section id="annotations">
<h2>Annotation Categories</h2>
<!-- Dataset graph toggle -->
<div style="display:flex;gap:0.4rem;margin-bottom:1rem;">
<button class="btn-pill btn-dark" id="graph-btn-covla" onclick="showGraph('covla')" style="font-size:0.82rem;padding:0.4rem 1rem;">CoVLA</button>
<button class="btn-pill" id="graph-btn-hsd" onclick="showGraph('hsd')" style="font-size:0.82rem;padding:0.4rem 1rem;background:#e8e8e8;color:#555;">Honda Scenes</button>
</div>
<div class="figure" id="graph-covla">
<img src="assets/phrase_graph_new.png" alt="CoVLA Category Visualization">
<div class="caption">Visualization of all CoVLA textual descriptions automatically categorized by VENUSS. The framework identified seven categories: motion types (blue), velocity descriptors (yellow), directional behaviors (orange), acceleration patterns (purple), following behavior (dark green), traffic light conditions (light red), and road curvature detection (gray).</div>
</div>
<div class="figure" id="graph-hsd" style="display:none;">
<img src="assets/hsd_phrase_graph.png" alt="HSD Category Visualization">
<div class="caption">Visualization of all Honda Scenes textual descriptions automatically categorized by VENUSS. The framework identified six categories: temporal states (blue), place types (orange), road types (gold), weather conditions (plum), surface conditions (coral), and ambient lighting (green). Gray connector nodes show the grammatical structure. Dashed edges indicate sentence termination points.</div>
</div>
<script>
function showGraph(ds) {
document.getElementById('graph-covla').style.display = ds === 'covla' ? '' : 'none';
document.getElementById('graph-hsd').style.display = ds === 'hsd' ? '' : 'none';
var btnC = document.getElementById('graph-btn-covla');
var btnH = document.getElementById('graph-btn-hsd');
if (ds === 'covla') {
btnC.className = 'btn-pill btn-dark';
btnH.className = 'btn-pill';
btnH.style.background = '#e8e8e8'; btnH.style.color = '#555';
btnC.style.background = ''; btnC.style.color = '';
} else {
btnH.className = 'btn-pill btn-dark';
btnC.className = 'btn-pill';
btnC.style.background = '#e8e8e8'; btnC.style.color = '#555';
btnH.style.background = ''; btnH.style.color = '';
}
}
</script>
<h3>Human Evaluation</h3>
<div class="figure">
<img src="assets/web_app_simple.png" alt="Human evaluation web application" style="max-width:700px;">
<div class="caption">Simplified example of the human evaluation questionnaire. The interface presents sequential driving images with temporal intervals, and seven multiple-choice questions for the corresponding categories. The final answer key concatenates the responses for comparison with ground truth annotations.</div>
</div>
<h3>Semantic Mapping Example</h3>
<p>The annotation system transforms natural language descriptions into structured answer keys. For example:</p>
<pre><code>Caption: "The ego vehicle is moving straight at a high speed"
+-------------------- Motion: Moving (A)
| +----------------- Direction: Straight (A)
| | +-------------- Speed: High speed (B)
| | | +----------- Following: Following (A)
| | | | +-------- Acceleration: Constant Speed (C)
| | | | | +----- Traffic Light: No traffic light (B)
| | | | | | +-- Curve: No curve (B)
| | | | | | |
Answer key: A A B A C B B</code></pre>
</section>
<!-- ============================================================ -->
<!-- KEY RESULTS -->
<!-- ============================================================ -->
<section id="results">
<h2>Key Results</h2>
<div class="card-grid">
<div class="stat-card">
<div class="number">57%</div>
<div class="label">Best VLM Accuracy<br>(Qwen-VL-Max)</div>
</div>
<div class="stat-card">
<div class="number">65%</div>
<div class="label">Human Baseline<br>(GIF mode)</div>
</div>
<div class="stat-card">
<div class="number">25+</div>
<div class="label">VLMs<br>Evaluated</div>
</div>
<div class="stat-card">
<div class="number">2,600+</div>
<div class="label">Scenarios<br>Tested</div>
</div>
</div>
<h3>Overall Performance</h3>
<p>Model family hierarchy (accuracy): <strong>Qwen (52.8%)</strong> > Gemini (51.2%) > Claude (50.5%) > GPT (49.4%). Qwen architectures demonstrate both superior accuracy and consistency (σ=0.032) compared to GPT models (σ=0.082), indicating 2.56× better reliability.</p>
<p>GPT-4o-Mini outperforms the flagship GPT-4o by 13 percentage points (55% vs 42%) due to GPT-4o's frequent refusals on driving-related content. Gemini-2.5-Pro was excluded entirely (90%+ refusal rate).</p>
<h3>Task Difficulty Hierarchy</h3>
<div class="table-wrapper">
<table>
<thead>
<tr>
<th>Task</th>
<th>Difficulty</th>
<th>F1 Range</th>
</tr>
</thead>
<tbody>
<tr><td style="text-align:left;">Vehicle Motion Detection</td><td><span class="badge badge-easy">Easy</span></td><td>0.73 – 0.91</td></tr>
<tr><td style="text-align:left;">Traffic Light Detection</td><td><span class="badge badge-moderate">Moderate</span></td><td>0.61 – 0.90</td></tr>
<tr><td style="text-align:left;">Curved Road Detection</td><td><span class="badge badge-moderate">Moderate-Hard</span></td><td>0.40 – 0.73</td></tr>
<tr><td style="text-align:left;">Car Following Behavior</td><td><span class="badge badge-hard">Hard</span></td><td>0.29 – 0.64</td></tr>
<tr><td style="text-align:left;">Vehicle Direction Analysis</td><td><span class="badge badge-hard">Hard</span></td><td>0.08 – 0.33</td></tr>
<tr><td style="text-align:left;">Vehicle Speed Assessment</td><td><span class="badge badge-hard">Very Hard</span></td><td>0.12 – 0.32</td></tr>
<tr><td style="text-align:left;">Vehicle Acceleration Detection</td><td><span class="badge badge-hard">Hardest</span></td><td>0.07 – 0.25</td></tr>
</tbody>
</table>
</div>
</section>
<!-- ============================================================ -->
<!-- CONFIGURATION ANALYSIS -->
<!-- ============================================================ -->
<section id="configuration">
<h2>Configuration Analysis</h2>
<p>Our systematic evaluation across <strong>270 unique configuration combinations</strong> reveals a <strong>48.2% relative improvement potential</strong> through systematic configuration tuning.</p>
<h3>Key Findings</h3>
<details>
<summary>Resolution: 720p is optimal</summary>
<div class="details-content">
<p>720p (960×540) achieves <strong>95% of peak performance</strong> while maintaining computational tractability. Lower resolutions lose critical visual details needed for scene interpretation, while higher resolutions add pixel density without new semantic information for these tasks.</p>
<div class="figure">
<img src="assets/resolution_comparison_with_gif_humans_box.png" alt="Resolution comparison" style="max-width:600px;">
</div>
</div>
</details>
<details>
<summary>Temporal Intervals: 1000ms intervals are best</summary>
<div class="details-content">
<p>Optimal performance at <strong>1000ms intervals (0.5918 accuracy)</strong> versus 200ms intervals (0.5068 accuracy), a <strong>14.4% accuracy penalty</strong> for shorter intervals. Longer intervals capture more visually distinct frames, making scene changes more apparent.</p>
<div class="figure">
<img src="assets/time_interval_comparison_with_gif_humans_box.png" alt="Time interval comparison" style="max-width:600px;">
</div>
</div>
</details>
<details>
<summary>Frame Count: 3-4 frames optimal, plateaus beyond</summary>
<div class="details-content">
<p>Increasing frames from 1 to 3-4 yields a <strong>31.5% improvement</strong>, but performance plateaus beyond 4 frames. Current VLMs cannot effectively integrate information across long sequences, likely due to attention mechanisms that struggle to track temporal changes across many frames.</p>
<div class="figure">
<img src="assets/num_images_comparison_with_gif_humans_box.png" alt="Number of images comparison" style="max-width:600px;">
</div>
</div>
</details>
<details>
<summary>Presentation Mode: Collage outperforms sequential by 6%</summary>
<div class="details-content">
<p>Spatial co-location allows models to compare frames simultaneously within a single image, rather than relying on context memory across separate inputs.</p>
<div class="figure">
<img src="assets/presentation_mode_comparison_with_gif_humans_box.png" alt="Presentation mode comparison" style="max-width:600px;">
</div>
</div>
</details>
<details>
<summary>Spatial Layout: Horizontal layouts beat square grids</summary>
<div class="details-content">
<p>Horizontal layouts (3×1, 4×1) outperform square grids (p < 0.001, η² = 0.41). Left-to-right arrangement matches natural temporal flow and avoids ambiguity in reading order.</p>
<div class="figure">
<img src="assets/grid_format_comparison_with_gif_humans_box.png" alt="Grid format comparison" style="max-width:600px;">
</div>
</div>
</details>
<!-- Interactive Configuration Explorer -->
<h3 style="margin-top:2rem;">Interactive Configuration Explorer</h3>
<p>Select a model to highlight its performance trend across all five configuration dimensions. The connecting line reveals how each model responds to different settings.</p>
<div id="interactive-explorer">
<div class="explorer-controls">
<label for="model-select"><strong>Highlight Model:</strong></label>
<select id="model-select"><option value="">Loading...</option></select>
</div>
<div class="explorer-legend" id="explorer-legend"></div>
<div class="explorer-grid">
<div id="plot-resolution" class="explorer-plot"></div>
<div id="plot-time_interval" class="explorer-plot"></div>
<div id="plot-num_images" class="explorer-plot"></div>
<div id="plot-presentation_mode" class="explorer-plot"></div>
</div>
<div id="plot-grid_format" class="explorer-plot explorer-plot-wide"></div>
</div>
<style>
#interactive-explorer { margin-top: 1rem; }
.explorer-controls {
display: flex; align-items: center; gap: 0.75rem;
margin-bottom: 0.75rem; flex-wrap: wrap;
}
.explorer-controls select {
padding: 0.4rem 0.6rem; font-size: 0.95rem;
border: 1px solid #ccc; border-radius: 4px;
min-width: 260px; font-family: inherit;
}
.explorer-legend {
display: flex; flex-wrap: wrap; gap: 1rem;
margin-bottom: 1rem; font-size: 0.85rem;
}
.explorer-legend span {
display: inline-flex; align-items: center; gap: 0.3rem;
}
.legend-swatch {
display: inline-block; width: 14px; height: 14px;
border-radius: 3px; border: 1px solid #999;
}
.legend-line {
display: inline-block; width: 20px; height: 0;
border-top: 3px solid; vertical-align: middle;
}
.explorer-grid {
display: grid;
grid-template-columns: 1fr 1fr;
gap: 0.5rem;
}
.explorer-plot { width: 100%; }
.explorer-plot-wide { margin-top: 0.5rem; width: 100%; }
@media (max-width: 768px) {
.explorer-grid { grid-template-columns: 1fr; }
}
</style>
<script>
(function() {
const PLOT_IDS = ['resolution', 'time_interval', 'num_images', 'presentation_mode', 'grid_format'];
const PLOT_HEIGHT = 340;
const GRID_PLOT_HEIGHT = 380;
const HUMAN_COLLAGE_COLOR = '#9B59B6';
const HUMAN_GIF_COLOR = '#E67E22';
let DATA = null;
fetch('assets/interactive_data.json')
.then(r => r.json())
.then(data => { DATA = data; init(); })
.catch(err => {
document.getElementById('model-select').innerHTML = '<option>Error loading data</option>';
console.error(err);
});
function init() {
buildDropdown();
buildLegend();
PLOT_IDS.forEach(renderPlot);
}
function buildDropdown() {
const sel = document.getElementById('model-select');
sel.innerHTML = '<option value="">-- None (show all) --</option>';
const byFamily = {};
for (const [name, info] of Object.entries(DATA.models)) {
(byFamily[info.family] = byFamily[info.family] || []).push({name, label: info.label});
}
for (const fam of ['claude', 'gemini', 'gpt', 'qwen']) {
if (!byFamily[fam]) continue;
const og = document.createElement('optgroup');
og.label = fam.charAt(0).toUpperCase() + fam.slice(1);
byFamily[fam].sort((a,b) => a.label.localeCompare(b.label)).forEach(m => {
const opt = document.createElement('option');
opt.value = m.name; opt.textContent = m.label;
og.appendChild(opt);
});
sel.appendChild(og);
}
sel.addEventListener('change', () => PLOT_IDS.forEach(updatePlot));
}
function buildLegend() {
const el = document.getElementById('explorer-legend');
let html = '';
for (const [fam, info] of Object.entries(DATA.families)) {
html += `<span><span class="legend-swatch" style="background:${info.color}"></span>${fam.charAt(0).toUpperCase() + fam.slice(1)}</span>`;
}
html += `<span><span class="legend-line" style="border-color:${HUMAN_COLLAGE_COLOR}"></span>Human (Collage)</span>`;
html += `<span><span class="legend-line" style="border-color:${HUMAN_GIF_COLOR};border-style:dashed"></span>Human (GIF)</span>`;
el.innerHTML = html;
}
function renderPlot(plotKey) {
const plot = DATA.plots[plotKey];
const divId = 'plot-' + plotKey;
const traces = [];
// VLM scatter traces
for (const [model, vals] of Object.entries(plot.vlm_data)) {
const info = DATA.models[model];
if (!info) continue;
// Filter nulls
const xs = [], ys = [];
vals.forEach((v, i) => { if (v !== null) { xs.push(plot.xvalues[i]); ys.push(v); } });
traces.push({
x: xs, y: ys,
mode: 'markers',
type: 'scatter',
name: info.label,
text: xs.map(() => info.label),
hovertemplate: '%{text}<br>%{x}: %{y:.4f}<extra></extra>',
marker: {
symbol: info.marker,
size: 8,
color: info.color,
line: { color: '#333', width: 0.5 },
opacity: 0.8,
},
meta: { modelName: model },
showlegend: false,
});
}
// Human collage baseline
if (plot.human_collage.some(v => v !== null)) {
const xs = [], ys = [];
plot.human_collage.forEach((v, i) => {
if (v !== null) { xs.push(plot.xvalues[i]); ys.push(v); }
});
traces.push({
x: xs, y: ys,
mode: 'lines',
type: 'scatter',
name: 'Human (Collage)',
line: { color: HUMAN_COLLAGE_COLOR, width: 3, dash: 'solid' },
opacity: 0.7,
hovertemplate: 'Human (Collage)<br>%{x}: %{y:.4f}<extra></extra>',
showlegend: false,
});
}
// Human GIF baseline
if (plot.human_gif.some(v => v !== null)) {
const xs = [], ys = [];
plot.human_gif.forEach((v, i) => {
if (v !== null) { xs.push(plot.xvalues[i]); ys.push(v); }
});
traces.push({
x: xs, y: ys,
mode: 'lines',
type: 'scatter',
name: 'Human (GIF)',
line: { color: HUMAN_GIF_COLOR, width: 3, dash: 'dash' },
opacity: 0.7,
hovertemplate: 'Human (GIF)<br>%{x}: %{y:.4f}<extra></extra>',
showlegend: false,
});
}
const isGrid = plotKey === 'grid_format';
const layout = {
title: { text: plot.title, font: { size: 15 } },
xaxis: {
title: plot.xlabel,
tickvals: plot.xvalues,
ticktext: plot.xlabels,
tickangle: isGrid ? -45 : 0,
tickfont: { size: isGrid ? 9 : 11 },
},
yaxis: { title: 'Accuracy', rangemode: 'tozero' },
height: isGrid ? GRID_PLOT_HEIGHT : PLOT_HEIGHT,
margin: { t: 40, b: isGrid ? 80 : 50, l: 55, r: 15 },
hovermode: 'closest',
plot_bgcolor: '#fafafa',
};
Plotly.newPlot(divId, traces, layout, { responsive: true, displayModeBar: false });
}
function updatePlot(plotKey) {
const selected = document.getElementById('model-select').value;
const divId = 'plot-' + plotKey;
const div = document.getElementById(divId);
const plot = DATA.plots[plotKey];
const update = { 'marker.opacity': [], 'marker.size': [] };
const traceIndices = [];
let highlightX = [], highlightY = [];
let highlightColor = '#333';
let traceIdx = 0;
for (const [model, vals] of Object.entries(plot.vlm_data)) {
const info = DATA.models[model];
if (!info) continue;
if (selected && model === selected) {
update['marker.opacity'].push(1.0);
update['marker.size'].push(12);
highlightColor = info.color;
// Collect points for line
vals.forEach((v, i) => {
if (v !== null) { highlightX.push(plot.xvalues[i]); highlightY.push(v); }
});
} else if (selected) {
update['marker.opacity'].push(0.15);
update['marker.size'].push(6);
} else {
update['marker.opacity'].push(0.8);
update['marker.size'].push(8);
}
traceIndices.push(traceIdx);
traceIdx++;
}
Plotly.restyle(divId, update, traceIndices);
// Remove old highlight line if present, then add new one
const existingTraces = div.data.length;
// Count non-VLM traces (human baselines)
const vlmCount = Object.keys(plot.vlm_data).filter(m => DATA.models[m]).length;
const baselineCount = existingTraces - vlmCount;
// Highlight line is any trace beyond VLM + baselines
const expectedBase = vlmCount + (plot.human_collage.some(v => v !== null) ? 1 : 0) +
(plot.human_gif.some(v => v !== null) ? 1 : 0);
if (existingTraces > expectedBase) {
Plotly.deleteTraces(divId, Array.from({length: existingTraces - expectedBase}, (_, i) => expectedBase + i));
}
if (selected && highlightX.length > 1) {
Plotly.addTraces(divId, {
x: highlightX, y: highlightY,
mode: 'lines',
type: 'scatter',
line: { color: highlightColor, width: 2.5 },
opacity: 0.9,
hoverinfo: 'skip',
showlegend: false,
});
}
}
})();
</script>
</section>
<!-- ============================================================ -->
<!-- HUMAN EVALUATION -->
<!-- ============================================================ -->
<section id="human-eval">
<h2>Human Evaluation</h2>
<h3>Evaluation Protocol</h3>
<p>The web application supports three presentation modes for human evaluators:</p>
<ol>
<li><strong>Image collages</strong> – matching the VLM input format for direct performance comparison</li>
<li><strong>Animated GIFs</strong> – providing temporal continuity for more intuitive scenario understanding</li>
<li><strong>Video playback</strong> – for annotation verification and data curation</li>
</ol>
<p>8 evaluators participated: 3 using collages, 5 using GIFs (108 evaluations each). GIF-based evaluations consistently outperformed collage-based (65% vs 56% peak accuracy, +16% relative improvement).</p>
<h3>Inter-Annotator Agreement (Fleiss' Kappa)</h3>
<div class="table-wrapper">
<table>
<thead>
<tr>
<th>Category</th>
<th>κ GIF (5 raters)</th>
<th>κ Collage (3 raters)</th>
</tr>
</thead>
<tbody>
<tr><td style="text-align:left;">Motion</td><td>0.595 (Moderate)</td><td>0.465 (Moderate)</td></tr>
<tr><td style="text-align:left;">Direction</td><td>0.815 (Almost perfect)</td><td>0.763 (Substantial)</td></tr>
<tr><td style="text-align:left;">Speed</td><td>0.672 (Substantial)</td><td>0.516 (Moderate)</td></tr>
<tr><td style="text-align:left;">Following</td><td>0.919 (Almost perfect)</td><td>0.907 (Almost perfect)</td></tr>
<tr><td style="text-align:left;">Acceleration</td><td>0.432 (Moderate)</td><td>0.248 (Fair)</td></tr>
<tr><td style="text-align:left;">Traffic Light</td><td>0.819 (Almost perfect)</td><td>0.788 (Substantial)</td></tr>
<tr><td style="text-align:left;">Curve</td><td>0.639 (Substantial)</td><td>0.585 (Moderate)</td></tr>
<tr style="font-weight:600; background:#f0f7ff;"><td style="text-align:left;">Overall Mean</td><td>0.699 (Substantial)</td><td>0.610 (Substantial)</td></tr>
</tbody>
</table>
</div>
<p>Both evaluation modes show substantial agreement. GIF consistently outperforms collage (κ +0.05–0.18 per category). Categories dependent on temporal cues (speed, motion, acceleration) show the largest improvement from collage to GIF, while spatially-grounded categories (following, traffic light) remain stable.</p>
</section>
<!-- ============================================================ -->
<!-- SUPPLEMENTARY MATERIAL -->
<!-- ============================================================ -->
<section id="supplementary">
<h2><span class="supp-badge">SUPPLEMENTARY</span> Prompt Templates</h2>
<p>The following sections present the complete prompt templates used for VLM evaluation in VENUSS. These could not be included in the main paper due to space constraints.</p>
<details>
<summary>System Prompt</summary>
<div class="details-content">
<p class="code-label">System Message</p>
<pre><code>You are an expert in autonomous driving scenario analysis. You will be shown images
representing sequential driving scenarios and must classify them according to specific
categories.</code></pre>
</div>
</details>
<details>
<summary>User Prompt – CoVLA Dataset</summary>
<div class="details-content">
<p>The user prompt provides context about the image content and task requirements. For each evaluation, models receive:</p>
<p class="code-label">User Message</p>
<pre><code>You are viewing a [rows]x[cols] grid of images showing a driving scenario captured
at [interval] ms intervals. The images are arranged chronologically from left to
right, top to bottom. These images are from the ego vehicle's perspective in a
driving scenario.
Based on the sequential images, please answer the following questions:
1. Motion State: Is the ego vehicle moving, stopping, or stopped?
Options: (A) Moving (B) Stopping (C) Stopped
2. Direction: What direction is the ego vehicle moving?
Options: (A) Straight (B) Left (C) Right
3. Speed: What is the ego vehicle's speed?
Options: (A) Very High (B) High (C) Medium (D) Low
4. Following: Is the ego vehicle following another vehicle?
Options: (A) Not following (B) Following
5. Acceleration: What is the ego vehicle's acceleration state?
Options: (A) Positive acceleration (B) Negative acceleration (C) No acceleration
6. Traffic Light: Is there a traffic light visible?
Options: (A) Red (B) No traffic light (C) Green (D) Yellow
7. Curve: Is the ego vehicle on a curved road?
Options: (A) Curved road (B) Straight road
Please provide your analysis and then include a short answer key at the end
in the format: 1) [letter] 2) [letter] ... 7) [letter]</code></pre>
<p><strong>Note:</strong> Models generate free-form responses and are free to express themselves. The answer key at the end is parsed for evaluation, enabling reproducible and scalable assessment across 25+ models and 2,600+ scenarios.</p>
</div>
</details>
<details>
<summary>User Prompt – Honda Scenes Dataset</summary>
<div class="details-content">
<p>Honda Scenes uses 16 environmental categories instead of CoVLA's 7 behavioral categories, covering road type, weather, surface conditions, ambient lighting, and various infrastructure elements.</p>
<p class="code-label">User Message</p>
<pre><code>You are viewing a [rows]x[cols] grid of images showing a driving scenario captured
at [interval] ms intervals. The images are arranged chronologically from left to
right, top to bottom. These images are from the ego vehicle's perspective in a
driving scenario.
Based on the sequential images, please answer the following questions about the
driving environment:
1. Road Type: What type of road environment is this?
Options: (A) Local (B) Highway (C) Ramp (D) Urban (E) Rural (F) NDA
2. Weather: What are the current weather conditions?
Options: (A) Sunny (B) Cloudy (C) Rain (D) Snow (E) Fog (F) Hail (G) NDA
3. Road Surface: What is the road surface condition?
Options: (A) Dry (B) Wet (C) Icy (D) Snow (E) NDA
4. Ambient Lighting: What is the lighting/time of day?
Options: (A) Day (B) Dawn/Dusk (C) Night (D) NDA
5. Intersection: Is there an intersection present in this scene?
Options: (A) Yes, 3-way (B) Yes, 4-way (C) Yes, 5+ way (D) No intersection (E) NDA
6. Traffic Control: What type of traffic control exists at the intersection (if present)?
Options: (A) Traffic lights (B) Stop sign (C) No signal (D) No intersection (E) NDA
7. Merge/Branch: Is there a merge or branch scenario visible?
Options: (A) Merge - gore on left (B) Merge - gore on right
(C) Branch - gore on left (D) Branch - gore on right
(E) No merge/branch (F) NDA
8. Bridge/Overpass: Is there a bridge or overpass in this scene?
Options: (A) Yes - overhead bridge (B) Yes - under overpass (C) No (D) NDA
9. Railway Crossing: Is there a railway crossing visible?
Options: (A) Yes (B) No (C) NDA
10. Tunnel: Is there a tunnel in this scene?
Options: (A) Yes (B) No (C) NDA
11. Construction Zone: Is there a construction zone visible?
Options: (A) Yes (B) No (C) NDA
12. Zebra Crossing: Is there a zebra crossing (marked pedestrian crossing) visible?
Options: (A) Yes (B) No (C) NDA
13. Regular Crosswalk: Is there a regular crosswalk visible?
Options: (A) Yes (B) No (C) NDA
14. Driveway: Is there a driveway entrance/exit visible?
Options: (A) Yes (B) No (C) NDA
15. Primary Scenario: What is the primary road scenario in this scene?
Options: (A) Intersection (B) Merge/Branch (C) Railway crossing
(D) Tunnel (E) Bridge/Overpass (F) Construction zone