Repository navigation
Expand file tree
/
Copy pathindex.html
More file actions
1028 lines (937 loc) · 44.2 KB
/
Copy pathindex.html
File metadata and controls
1028 lines (937 loc) · 44.2 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>
<head>
<meta charset="utf-8">
<meta name="description"
content="dToF Depth Completion — CVPR 2026 project page.">
<meta name="keywords" content="dToF, Depth Completion, CVPR 2026">
<meta name="viewport" content="width=device-width, initial-scale=1">
<title>dToF Depth Completion — CVPR 2026</title>
<link href="https://fonts.googleapis.com/css?family=Google+Sans|Noto+Sans|Castoro"
rel="stylesheet">
<link rel="stylesheet" href="./static/css/bulma.min.css">
<link rel="stylesheet" href="./static/css/bulma-carousel.min.css">
<link rel="stylesheet" href="./static/css/bulma-slider.min.css">
<link rel="stylesheet" href="./static/css/fontawesome.all.min.css">
<link rel="stylesheet"
href="https://cdn.jsdelivr.net/gh/jpswalsh/academicons@1/css/academicons.min.css">
<link rel="stylesheet" href="./static/css/index.css">
<link rel="icon" href="./static/images/favicon.svg">
<style>
/* ===== CVPR 2026 Badge ===== */
.cvpr-badge {
display: inline-block;
margin-bottom: 18px;
padding: 6px 18px;
border-radius: 100px;
background: #2563eb;
color: #fff;
font-size: 15px;
font-weight: 700;
letter-spacing: 1px;
text-transform: uppercase;
}
/* ===== Results Grid ===== */
.results-grid {
display: grid;
grid-template-columns: repeat(auto-fit, minmax(280px, 1fr));
gap: 20px;
margin-top: 28px;
}
.result-card {
border-radius: 10px;
overflow: hidden;
border: 1px solid #e5e7eb;
transition: box-shadow 0.2s;
}
.result-card:hover { box-shadow: 0 6px 20px rgba(0,0,0,0.08); }
.result-card img, .result-card video { width: 100%; display: block; }
.result-card .caption {
padding: 12px 16px;
font-size: 13px;
color: #555;
background: #f8fafc;
}
/* ===== Results Triple Slider ===== */
.triple-slider {
position: relative; max-width: 900px; margin: 28px auto 0;
border-radius: 10px; overflow: hidden;
cursor: col-resize; user-select: none; -webkit-user-select: none;
border: 1px solid #e5e7eb;
}
.triple-slider .base { display: block; width: 100%; pointer-events: none; }
.triple-slider .layer {
position: absolute; top: 0; left: 0; width: 100%; height: 100%;
pointer-events: none;
}
.triple-slider .layer img {
position: absolute; top: 0; left: 0;
width: 100%; height: 100%; object-fit: cover;
}
.triple-slider .ts-divider {
position: absolute; top: 0; width: 3px; height: 100%;
background: #fff; box-shadow: 0 0 8px rgba(0,0,0,0.4);
z-index: 5; transform: translateX(-50%);
}
.triple-slider .ts-handle {
position: absolute; top: 50%; left: 0;
transform: translate(-50%, -50%);
width: 34px; height: 34px;
background: #fff; border-radius: 50%;
box-shadow: 0 2px 10px rgba(0,0,0,0.25);
z-index: 6; cursor: col-resize;
display: flex; align-items: center; justify-content: center;
}
.triple-slider .ts-handle::before {
content: '\25C0 \25B6'; font-size: 8px; color: #555; letter-spacing: 2px;
}
.triple-slider .ts-label {
position: absolute; bottom: 12px;
padding: 4px 14px; border-radius: 6px;
font-size: 12px; font-weight: 600; color: #fff;
z-index: 4; pointer-events: none; backdrop-filter: blur(6px);
}
.triple-slider .ts-label.rgb { left: 12px; background: rgba(55, 65, 81, 0.85); }
.triple-slider .ts-label.pred { left: 50%; transform: translateX(-50%); background: rgba(37, 99, 235, 0.85); }
.triple-slider .ts-label.sparse { right: 12px; background: rgba(30, 41, 59, 0.85); }
/* ===== Comparison Slider ===== */
.comp-scene-selector {
display: flex; flex-wrap: wrap; justify-content: center;
gap: 8px; margin-bottom: 24px;
}
.comp-scene-btn {
padding: 7px 18px; border-radius: 100px;
border: 1px solid #e5e7eb; background: #fff;
font-size: 13px; font-weight: 500; cursor: pointer;
color: #555; transition: all 0.2s;
}
.comp-scene-btn:hover { border-color: #2563eb; color: #2563eb; }
.comp-scene-btn.active {
background: #2563eb; color: #fff; border-color: #2563eb;
}
.comp-slider {
position: relative; max-width: 800px; margin: 0 auto;
border-radius: 10px; overflow: hidden;
cursor: col-resize; user-select: none; -webkit-user-select: none;
border: 1px solid #e5e7eb;
}
.comp-slider img { display: block; width: 100%; pointer-events: none; }
.comp-slider .overlay {
position: absolute; top: 0; left: 0; height: 100%; overflow: hidden;
}
.comp-slider .overlay img {
display: block; height: 100%; width: auto;
min-width: 0; max-width: none; object-fit: cover;
}
.comp-slider .divider {
position: absolute; top: 0; width: 3px; height: 100%;
background: #fff; box-shadow: 0 0 8px rgba(0,0,0,0.4);
z-index: 2; pointer-events: none;
}
.comp-slider .handle {
position: absolute; top: 50%;
transform: translate(-50%, -50%);
width: 36px; height: 36px;
background: #fff; border-radius: 50%;
box-shadow: 0 2px 10px rgba(0,0,0,0.25);
z-index: 3; pointer-events: none;
display: flex; align-items: center; justify-content: center;
}
.comp-slider .handle::before {
content: '\25C0 \25B6'; font-size: 9px; color: #555; letter-spacing: 2px;
}
.comp-slider .label-left, .comp-slider .label-right {
position: absolute; bottom: 12px;
padding: 4px 14px; border-radius: 6px;
font-size: 12px; font-weight: 600;
z-index: 4; pointer-events: none;
backdrop-filter: blur(6px);
}
.comp-slider .label-left {
left: 12px; background: rgba(37, 99, 235, 0.85); color: #fff;
}
.comp-slider .label-right {
right: 12px; background: rgba(225, 29, 72, 0.85); color: #fff;
}
/* ===== Comparison layout: inputs (col 1) | divider | results (cols 3-4) =====
One unified grid so every input item and every result panel share the SAME
column width — and therefore the same height — keeping the input column
flush (top and bottom) with the results column. */
.comp-layout {
display: grid;
grid-template-columns: 1fr 1px 1fr 1fr; /* inputs | divider | result | result */
grid-template-rows: repeat(3, auto);
column-gap: 24px;
row-gap: 16px;
align-items: start;
}
.cmp-inputs, .pc-grid { display: contents; } /* children flow into .comp-layout */
.comp-divider { /* full-height vertical divider */
grid-column: 2; grid-row: 1 / -1;
align-self: stretch; background: #e5e7eb;
}
/* Inputs fill column 1 top-to-bottom; the six results auto-flow row-major
into columns 3-4 (GT top-left … Ours bottom-right). */
.cmp-input-item.is-rgb { grid-column: 1; grid-row: 1; }
.cmp-input-item.is-sensor { grid-column: 1; grid-row: 2; }
.pc-panel.is-input { grid-column: 1; grid-row: 3; }
/* ===== Point Cloud Viewers ===== */
@media (max-width: 768px) {
/* Stop the unified grid; stack inputs (row of 3) above results (2 cols). */
.comp-layout { display: block; }
.cmp-inputs { display: grid; grid-template-columns: repeat(3, 1fr); gap: 16px; margin-bottom: 16px; }
.pc-grid { display: grid; grid-template-columns: repeat(2, 1fr); gap: 16px; }
.comp-divider { display: none; }
.cmp-input-item.is-rgb, .cmp-input-item.is-sensor,
.pc-panel.is-input { grid-column: auto; grid-row: auto; }
}
@media (max-width: 480px) {
.cmp-inputs, .pc-grid { grid-template-columns: 1fr; }
}
.pc-panel {
border: 1px solid #e5e7eb; border-radius: 10px; overflow: hidden;
background: #fff;
}
.pc-canvas {
position: relative; width: 100%; aspect-ratio: 4 / 3;
}
.pc-canvas canvas { display: block; width: 100%; height: 100%; }
.pc-spinner {
position: absolute; inset: 0;
display: flex; align-items: center; justify-content: center;
color: #94a3b8; font-size: 13px;
}
.pc-label {
text-align: center; padding: 9px;
font-size: 13px; font-weight: 600; color: #fff;
background: rgba(37, 99, 235, 0.9); /* methods: blue */
}
.pc-panel.is-input .pc-label { background: rgba(107, 114, 128, 0.92); } /* inputs: gray */
.pc-panel.is-ours .pc-label { background: rgba(22, 163, 74, 0.92); } /* our method: green */
.pc-panel.is-gt .pc-label { background: rgba(22, 163, 74, 0.92); } /* GT: green */
/* ===== Comparison Input Row (RGB + sensor depth) =====
Layout for .cmp-inputs lives in the .comp-layout rules above (it uses
display:contents on desktop and a stacked grid on mobile). */
.cmp-input-item {
border: 1px solid #e5e7eb; border-radius: 10px; overflow: hidden;
background: #f8fafc;
}
.cmp-input-item img {
display: block; width: 100%; aspect-ratio: 4 / 3; object-fit: cover;
}
.cmp-input-item.is-sensor img {
background: #0b1020;
image-rendering: pixelated; /* keep sparse dToF samples crisp */
}
.cmp-input-item figcaption {
text-align: center; padding: 9px;
font-size: 13px; font-weight: 600; color: #fff;
}
.cmp-input-item.is-rgb figcaption,
.cmp-input-item.is-sensor figcaption { background: rgba(37, 99, 235, 0.9); }
/* ===== More Results Triplets ===== */
.more-results-scene { margin-bottom: 36px; }
.more-results-scene h3 {
font-size: 15px; font-weight: 600;
color: #555; margin-bottom: 14px;
}
.result-triplet {
display: grid; grid-template-columns: repeat(3, 1fr); gap: 12px;
}
.result-triplet .result-item {
border-radius: 8px; overflow: hidden; border: 1px solid #e5e7eb;
}
.result-triplet .result-item img { width: 100%; display: block; }
.result-triplet .result-item .label {
text-align: center; padding: 8px;
font-size: 12px; font-weight: 500;
color: #555; background: #f8fafc;
}
@media (max-width: 640px) {
.result-triplet { grid-template-columns: 1fr; }
}
/* Teaser image (replaces the nerfies teaser video) */
.hero.teaser img { width: 100%; border-radius: 8px; }
</style>
<script src="https://ajax.googleapis.com/ajax/libs/jquery/3.5.1/jquery.min.js"></script>
<script defer src="./static/js/fontawesome.all.min.js"></script>
<script src="./static/js/bulma-carousel.min.js"></script>
<script src="./static/js/bulma-slider.min.js"></script>
<script src="./static/js/index.js"></script>
</head>
<body>
<section class="hero">
<div class="hero-body">
<div class="container is-max-desktop">
<div class="columns is-centered">
<div class="column has-text-centered">
<p class="cvpr-badge">CVPR 2026</p>
<h1 class="title is-1 publication-title">Dense Metric Depth Completion from Sparse Direct Time-of-Flight Sensors</h1>
<div class="is-size-5 publication-authors">
<span class="author-block">
<a href="https://sites.google.com/view/hakyeongkim">Hakyeong Kim</a><sup>1*</sup>,</span>
<span class="author-block">
<a href="https://wangrc.site/">Ruicheng Wang</a><sup>2, 3</sup>,</span>
<span class="author-block">
<a href="#">Chengtang Yao</a><sup>2</sup>,</span>
<span class="author-block">
<a href="https://jlyang.org/">Jiaolong Yang</a><sup>2</sup>,</span>
<span class="author-block">
<a href="https://vclab.kaist.ac.kr/minhkim">Min H. Kim</a><sup>1</sup></span>
</div>
<div class="is-size-5 publication-authors">
<span class="author-block"><sup>1</sup>KAIST,</span>
<span class="author-block"><sup>2</sup>Microsoft Research Asia</span>
<span class="author-block"><sup>3</sup>USTC</span>
</div>
<div class="column has-text-centered">
<div class="publication-links">
<span class="link-block">
<a href="./static/pdfs/dtof_dense_depth_main.pdf"
class="external-link button is-normal is-rounded is-dark">
<span class="icon"><i class="fas fa-file-pdf"></i></span>
<span>Paper</span>
</a>
</span>
<span class="link-block">
<a href="./static/pdfs/dtof_dense_depth_supp.pdf"
class="external-link button is-normal is-rounded is-dark">
<span class="icon"><i class="fas fa-file-pdf"></i></span>
<span>Supplement</span>
</a>
</span>
<span class="link-block">
<!-- <a href="#" class="external-link button is-normal is-rounded is-dark">
<span class="icon"><i class="ai ai-arxiv"></i></span>
<span>arXiv</span>
</a> -->
</span>
<span class="link-block">
<a href="https://www.youtube.com/watch?v=_0dIKkF1Oko" class="external-link button is-normal is-rounded is-dark">
<span class="icon"><i class="fab fa-youtube"></i></span>
<span>Video</span>
</a>
</span>
<span class="link-block">
<a href="https://github.com/KAIST-VCLAB/dToF-DMDC" class="external-link button is-normal is-rounded is-dark">
<span class="icon"><i class="fab fa-github"></i></span>
<span>Code</span>
</a>
</span>
</div>
</div>
</div>
</div>
</div>
</div>
</section>
<section class="hero teaser">
<div class="container is-max-desktop">
<div class="hero-body">
<img src="./static/images/teaser_user.jpg" alt="Teaser figure">
<h2 class="subtitle has-text-centered">
Zero-shot generalization of our model across different dToF sensing conditions. Top: sparse depth inputs from three representative settings—autonomous driving LiDAR (rotating dToF), lightweight mobile dToF, and extremely sparse random sampling. Bottom: our predicted dense metric depth for each case, demonstrating strong robustness under diverse sparsity, noise, and sensor patterns. Right: comparison of accuracy versus inference time (bubble size indicates model parameters). Our method achieves the most favorable balance of low error and fast runtime, outperforming state-of-the-art depth completion and enhancement approaches.
</h2>
</div>
</div>
</section>
<section class="section">
<div class="container is-max-desktop">
<!-- Abstract. -->
<div class="columns is-centered has-text-centered">
<div class="column is-four-fifths">
<h2 class="title is-3">Abstract</h2>
<div class="content has-text-justified">
<p>
Direct Time-of-Flight (dToF) sensors provide highly accurate metric depth and are more robust than indirect ToF systems in challenging real-world conditions. However, their high manufacturing cost and limited photodiode array size produce depth maps that are extremely sparse, low-resolution, and noisy, making them unsuitable for VR/XR, robotics, and 3D perception tasks that require dense metric depth. Existing monocular and depth completion methods struggle to handle the unique sampling patterns and hardware artifacts of dToF devices, and their performance often deteriorates significantly under severe sparsity or noise. We present a generalizable framework for dense metric depth completion from sparse dToF measurements, capable of operating across diverse sensor types, sparsity levels, and noise conditions. Our model employs a depth-guided dual-branch Vision Transformer encoder that processes RGB images and sparse dToF measurements separately, while a masked joint attention module allows depth tokens to reliably guide image features without being overwritten by them. A lightweight decoder reconstructs dense metric depth efficiently, without diffusion-based or refinement-heavy post-processing. To address the scarcity of paired training data, we introduce a comprehensive dToF simulation pipeline that reproduces the characteristics of flash, sub-VGA flash, and rotating sensors, including hardware-induced degradation, irregular sparsity, and realistic noise distributions. Trained entirely on synthetic data, our model achieves strong zero-shot generalization across 6 datasets and 3 real dToF devices, outperforming state-of-the-art approaches in both accuracy and computational efficiency. This establishes a robust and practical solution for dense metric depth completion from sparse direct ToF sensors.
</p>
</div>
</div>
</div>
</div>
</section>
<!-- ===================== RESULTS ===================== -->
<section class="section">
<div class="container is-max-desktop">
<h2 class="title is-3">Results</h2>
<p style="margin-bottom: 20px; color: #555;">
Pick a scene, then drag the two handles to reveal the RGB image, the sparse
depth input, and our completed depth within a single frame.
</p>
<div class="comp-scene-selector" id="tsSceneSelector">
<button class="comp-scene-btn active"
data-rgb="./static/images/main_results/KITTI_DC/30_2011_09_29_drive_0026__0000000086_image_02/rgb.jpg"
data-pred="static/images/main_results/KITTI_DC/30_2011_09_29_drive_0026__0000000086_image_02/pred_depth.png"
data-sparse="static/images/main_results/KITTI_DC/30_2011_09_29_drive_0026__0000000086_image_02/sensor.png"
>KITTI-DC 1</button>
<button class="comp-scene-btn active"
data-rgb="./static/images/main_results/KITTI_DC/03_2011_09_26_drive_0002__0000000023_image_02/rgb.jpg"
data-pred="static/images/main_results/KITTI_DC/03_2011_09_26_drive_0002__0000000023_image_02/pred_depth.png"
data-sparse="static/images/main_results/KITTI_DC/03_2011_09_26_drive_0002__0000000023_image_02/sensor.png"
>KITTI-DC 2</button>
<button class="comp-scene-btn"
data-rgb="./static/images/main_results/DIODE/20_val__indoors__scene_00020__scan_00185__00020_00185_indoors_050_010/rgb.jpg"
data-pred="static/images/main_results/DIODE/20_val__indoors__scene_00020__scan_00185__00020_00185_indoors_050_010/pred_depth.png"
data-sparse="static/images/main_results/DIODE/20_val__indoors__scene_00020__scan_00185__00020_00185_indoors_050_010/sensor.png"
>DIODE</button>
<button class="comp-scene-btn"
data-rgb="./static/images/main_results/DDAD/28_val__003453__CAMERA_05/rgb.jpg"
data-pred="static/images/main_results/DDAD/28_val__003453__CAMERA_05/pred_depth.png"
data-sparse="static/images/main_results/DDAD/28_val__003453__CAMERA_05/sensor.png"
>DDAD</button>
<button class="comp-scene-btn"
data-rgb="./static/images/main_results/ETH3D/44_electro__DSC_9302/rgb.jpg"
data-pred="static/images/main_results/ETH3D/44_electro__DSC_9302/pred_depth.png"
data-sparse="static/images/main_results/ETH3D/44_electro__DSC_9302/sensor.png"
>ETH3D</button>
<button class="comp-scene-btn"
data-rgb="./static/images/main_results/iBims/52_livingroom_04/rgb.jpg"
data-pred="static/images/main_results/iBims/52_livingroom_04/pred_depth.png"
data-sparse="static/images/main_results/iBims/52_livingroom_04/sensor.png"
>iBims 1</button>
<button class="comp-scene-btn"
data-rgb="./static/images/main_results/iBims/27_livingroom_03/rgb.jpg"
data-pred="static/images/main_results/iBims/27_livingroom_03/pred_depth.png"
data-sparse="static/images/main_results/iBims/27_livingroom_03/sensor.png"
>iBims 2</button>
<button class="comp-scene-btn"
data-rgb="./static/images/main_results/ZJUL5/67_cafe2__1646382413.602673/rgb.jpg"
data-pred="static/images/main_results/ZJUL5/67_cafe2__1646382413.602673/pred_depth.png"
data-sparse="static/images/main_results/ZJUL5/67_cafe2__1646382413.602673/sensor.png"
>ZJUL5</button>
</div>
<div class="triple-slider" id="tripleSlider">
<!-- base layer: RGB (left region, always full frame) -->
<img class="base" id="tsRGB" src="./static/images/main_results/KITTI_DC/30_2011_09_29_drive_0026__0000000086_image_02/rgb.jpg" alt="RGB Image">
<!-- middle layer: Predicted depth, revealed from divider 1 rightward -->
<div class="layer" id="tsPredLayer">
<img id="tsPred" src="static/images/main_results/KITTI_DC/30_2011_09_29_drive_0026__0000000086_image_02/pred_depth.png" alt="Predicted Depth">
</div>
<!-- top layer: Sparse depth, revealed from divider 2 rightward -->
<div class="layer" id="tsSparseLayer">
<img id="tsSparse" src="static/images/main_results/KITTI_DC/30_2011_09_29_drive_0026__0000000086_image_02/sensor.png" alt="Sparse Depth">
</div>
<div class="ts-divider" id="tsDiv1"></div>
<div class="ts-divider" id="tsDiv2"></div>
<div class="ts-handle" id="tsHandle1"></div>
<div class="ts-handle" id="tsHandle2"></div>
<span class="ts-label rgb">RGB</span>
<span class="ts-label pred">Predicted Depth</span>
<span class="ts-label sparse">Sparse Depth</span>
</div>
</div>
</section>
<!-- ===================== COMPARISON ===================== -->
<section class="section">
<div class="container is-max-desktop">
<h2 class="title is-3">Comparison</h2>
<p style="color: #555; margin-bottom: 20px;">
Interactive 3D point clouds. The sparse dToF input is shown alongside each
method’s completed result — all views share one synced camera, so dragging
to orbit (or scrolling to zoom) in any panel moves them all together.
</p>
<div class="comp-scene-selector" id="pcSceneSelector"></div>
<div class="comp-layout">
<!-- Left column: RGB input, sparse sensor depth, and the sparse point cloud
(the sparse-cloud viewer is injected into #cmpInputs by the script). -->
<div class="cmp-inputs" id="cmpInputs">
<figure class="cmp-input-item is-rgb">
<img id="cmpImage" src="" alt="Input image">
<figcaption>Input color image</figcaption>
</figure>
<figure class="cmp-input-item is-sensor">
<img id="cmpSensor" src="" alt="Input sensor depth">
<figcaption>Input depth map</figcaption>
</figure>
</div>
<!-- Vertical divider between inputs and results. -->
<div class="comp-divider" aria-hidden="true"></div>
<!-- Right column: comparison results — GT (top-left) … Ours (bottom-right). -->
<div class="pc-grid" id="pcGrid"></div>
</div>
<p style="color:#94a3b8; font-size:12px; text-align:center; margin-top:14px;">
Drag = rotate · Scroll = zoom · Right-drag = pan · all panels share one camera
</p>
</div>
</section>
<!-- ===================== VIDEO ===================== -->
<section class="section">
<div class="container is-max-desktop">
<div class="columns is-centered has-text-centered">
<div class="column is-four-fifths">
<h2 class="title is-3">Supplement Video</h2>
<div class="publication-video">
<iframe src="https://www.youtube.com/embed/f_XfNk0tqjs?rel=0&showinfo=0"
frameborder="0" allow="autoplay; encrypted-media" allowfullscreen></iframe>
</div>
</div>
</div>
</div>
</section>
<section class="section" id="BibTeX">
<div class="container is-max-desktop content">
<h2 class="title">BibTeX</h2>
<pre><code>
@InProceedings{Kim_2026_CVPR,
author = {Hakyeong Kim and Ruicheng Wang and
Chengtang Yao and Jiaolong Yang and Min H. Kim},
title = {Dense Metric Depth Completion from Sparse Direct Time-of-Flight Sensors},
booktitle = {IEEE Conference on Computer Vision and
Pattern Recognition (CVPR)},
month = {June},
year = {2026}
}
</code></pre>
</div>
</section>
<footer class="footer">
<div class="container">
<div class="columns is-centered">
<div class="column is-8">
<div class="content">
<p>
This website is licensed under a <a rel="license"
href="http://creativecommons.org/licenses/by-sa/4.0/">Creative
Commons Attribution-ShareAlike 4.0 International License</a>.
</p>
<p>
This website is adapted from the <a
href="https://github.com/nerfies/nerfies.github.io">Nerfies project page</a>.
</p>
</div>
</div>
</div>
</div>
</footer>
<!-- ===================== RESULTS TRIPLE SLIDER JS ===================== -->
<script>
(function () {
const slider = document.getElementById('tripleSlider');
const predLayer = document.getElementById('tsPredLayer');
const sparseLayer = document.getElementById('tsSparseLayer');
const div1 = document.getElementById('tsDiv1');
const div2 = document.getElementById('tsDiv2');
const handle1 = document.getElementById('tsHandle1');
const handle2 = document.getElementById('tsHandle2');
const MIN_GAP = 4; // keep the two dividers from crossing
let pos = [33, 66]; // divider positions (%)
let active = null; // which handle is being dragged
function render() {
// Predicted depth covers everything right of divider 1; sparse depth covers right of divider 2.
// => left region: RGB, middle: Predicted, right: Sparse.
predLayer.style.clipPath = 'inset(0 0 0 ' + pos[0] + '%)';
sparseLayer.style.clipPath = 'inset(0 0 0 ' + pos[1] + '%)';
div1.style.left = handle1.style.left = pos[0] + '%';
div2.style.left = handle2.style.left = pos[1] + '%';
}
function pctFromEvent(x) {
const rect = slider.getBoundingClientRect();
return Math.max(0, Math.min(100, ((x - rect.left) / rect.width) * 100));
}
function move(x) {
let p = pctFromEvent(x);
if (active === 0) pos[0] = Math.max(2, Math.min(p, pos[1] - MIN_GAP));
else pos[1] = Math.min(98, Math.max(p, pos[0] + MIN_GAP));
render();
}
// pick the nearest handle when pressing anywhere on the slider
function pickNearest(x) {
const p = pctFromEvent(x);
active = Math.abs(p - pos[0]) <= Math.abs(p - pos[1]) ? 0 : 1;
}
handle1.addEventListener('mousedown', (e) => { active = 0; e.preventDefault(); });
handle2.addEventListener('mousedown', (e) => { active = 1; e.preventDefault(); });
slider.addEventListener('mousedown', (e) => {
if (active === null) { pickNearest(e.clientX); move(e.clientX); }
});
window.addEventListener('mousemove', (e) => { if (active !== null) move(e.clientX); });
window.addEventListener('mouseup', () => { active = null; });
function touchStart(e, h) { active = h; }
handle1.addEventListener('touchstart', (e) => touchStart(e, 0), { passive: true });
handle2.addEventListener('touchstart', (e) => touchStart(e, 1), { passive: true });
slider.addEventListener('touchstart', (e) => {
if (active === null) { pickNearest(e.touches[0].clientX); move(e.touches[0].clientX); }
}, { passive: true });
window.addEventListener('touchmove', (e) => { if (active !== null) move(e.touches[0].clientX); }, { passive: true });
window.addEventListener('touchend', () => { active = null; });
// scene selector — swap the three images per scene
const rgbImg = document.getElementById('tsRGB');
const predImg = document.getElementById('tsPred');
const sparseImg = document.getElementById('tsSparse');
const tsButtons = document.querySelectorAll('#tsSceneSelector .comp-scene-btn');
tsButtons.forEach(btn => {
btn.addEventListener('click', () => {
tsButtons.forEach(b => b.classList.remove('active'));
btn.classList.add('active');
rgbImg.src = btn.dataset.rgb;
predImg.src = btn.dataset.pred;
sparseImg.src = btn.dataset.sparse;
});
});
render();
})();
</script>
<!-- ===================== POINT CLOUD VIEWERS (Three.js) ===================== -->
<script type="importmap">
{
"imports": {
"three": "./static/js/three/three.module.js",
"three/addons/": "./static/js/three/addons/"
}
}
</script>
<script type="module">
import * as THREE from 'three';
import { OrbitControls } from 'three/addons/controls/OrbitControls.js';
import { PLYLoader } from 'three/addons/loaders/PLYLoader.js';
/* ----------------------------------------------------------------------
EDIT HERE: one entry per scene. `dir` is a folder under PC_BASE that
holds input_image.jpg, input_sensor_depth_vis.png, and the .ply files.
Every scene must list the SAME panels in the same order (the first scene
defines the grid layout).
kind: 'input' -> dark label + enlarged points (sparse dToF)
kind: 'gt' -> green label
---------------------------------------------------------------------- */
const PC_BASE = 'static/images/comparisons/';
// Panel layout for every scene. The 'input' panel (sparse cloud) is rendered
// in the top inputs row; the rest fill the 2 x 3 grid in this order.
// Right-column grid fills row-major across 2 columns, so the order below maps to:
// GT | OmniDC
// Marigold | Prior-DA
// Prompt | Ours
// i.e. GT top-left, Ours bottom-right. ('input' is pulled out into the left column.)
const PANELS = () => ([
{ label: 'Input point cloud (visualization)', file: 'input_sensor_depth.ply', kind: 'input' },
{ label: 'GT depth', file: 'gt_depth.ply', kind: 'gt' },
{ label: 'OmniDC', file: 'omni_dc.ply' },
{ label: 'Marigold-DC', file: 'marigold_dc.ply' },
{ label: 'Prior-DA', file: 'prior_da.ply' },
{ label: 'Prompt-DA', file: 'prompt_da.ply' },
{ label: 'Ours', file: 'ours.ply', kind: 'ours' },
]);
const mkScene = (name, dir) => ({ name, dir, image: 'input_image.jpg', sensor: 'input_sensor_depth_vis.png', fov: 'fov.json', panels: PANELS() });
const PC_SCENES = [
mkScene('ZJUL5 — cafe', 'ZJUL5__cafe1_1645168915.171201'),
mkScene('iBims — restaurant', 'iBims-1_500__restaurant_11'),
{ ...mkScene('KITTI-DC — driving', 'KITTI_DC__2011_10_03_drive_0047_0000000602_image_03'), aspect: 16 / 9 },
mkScene('DIODE — indoors', 'DIODE_500__val_indoors_scene_00020_scan_00185_00020_00185_indoors_080_010'),
mkScene('ETH3D — relief', 'ETH3D_500__relief_DSC_0442'),
mkScene('ETH3D — terrains', 'ETH3D_500__terrains_DSC_0618'),
mkScene('DDAD — driving', 'DDAD_500__val_001989_CAMERA_01'),
];
const POINT_SIZE = 0.0058; // dense clouds: point size relative to framing radius
const SPARSE_SIZE = 0.026; // sparse dToF input: much larger so samples are visible
const FOV = 50; // fallback FOV when a scene has no fov.json
const REF_LABEL = 'Ours'; // all clouds are co-registered + framed to this one
// Fallback framing (used only when fov.json is missing): a gentle orbit off
// the frontal view so the 2.5D depth surface still reads as 3D.
const INIT_AZIM = THREE.MathUtils.degToRad(20);
const INIT_ELEV = THREE.MathUtils.degToRad(12);
const DIST_PAD = 1.15;
const grid = document.getElementById('pcGrid');
const inputsRow = document.getElementById('cmpInputs');
const selector = document.getElementById('pcSceneSelector');
const cmpImage = document.getElementById('cmpImage');
const cmpSensor = document.getElementById('cmpSensor');
const loader = new PLYLoader();
let viewers = [];
let syncing = false;
// A SINGLE shared WebGL context is rendered once per panel and blitted into
// that panel's lightweight 2D canvas. Creating one context (instead of one
// per panel) avoids the per-page WebGL context limit that makes context
// creation fail outright on some machines / GPU configurations.
const DPR = Math.min(window.devicePixelRatio || 1, 2);
let sharedRenderer = null;
function renderViewer(v) {
const w = v.canvas.width, h = v.canvas.height;
if (!w || !h || !sharedRenderer) return;
sharedRenderer.setSize(w, h, false);
sharedRenderer.render(v.scene, v.camera);
v.ctx.clearRect(0, 0, w, h);
v.ctx.drawImage(sharedRenderer.domElement, 0, 0, w, h);
}
function renderAll() {
viewers.forEach(renderViewer);
}
function resizeViewer(v) {
const w = v.holder.clientWidth, h = v.holder.clientHeight;
if (!w || !h) return;
v.canvas.width = Math.round(w * DPR);
v.canvas.height = Math.round(h * DPR);
v.camera.aspect = w / h;
v.camera.updateProjectionMatrix();
}
function makeViewer(def) {
const panel = document.createElement('div');
panel.className = 'pc-panel';
if (def.kind === 'input') panel.classList.add('is-input');
if (def.kind === 'gt') panel.classList.add('is-gt');
if (def.kind === 'ours') panel.classList.add('is-ours');
const holder = document.createElement('div');
holder.className = 'pc-canvas';
const spinner = document.createElement('div');
spinner.className = 'pc-spinner';
spinner.textContent = 'Loading…';
const labelEl = document.createElement('div');
labelEl.className = 'pc-label';
holder.appendChild(spinner);
panel.appendChild(holder);
panel.appendChild(labelEl);
// Sparse cloud sits in the top inputs row; everything else in the 2 x 3 grid.
(def.kind === 'input' ? inputsRow : grid).appendChild(panel);
const canvas = document.createElement('canvas');
holder.appendChild(canvas);
const ctx = canvas.getContext('2d');
const scene = new THREE.Scene();
scene.background = new THREE.Color(0xffffff);
const camera = new THREE.PerspectiveCamera(FOV, 1, 0.01, 1000);
camera.position.set(0, 0, 2.5);
const controls = new OrbitControls(camera, canvas);
controls.enableDamping = false;
controls.addEventListener('change', () => {
if (syncing) return;
syncing = true;
for (const o of viewers) {
if (o.camera !== camera) {
o.camera.position.copy(camera.position);
o.camera.quaternion.copy(camera.quaternion);
o.controls.target.copy(controls.target);
}
}
syncing = false;
renderAll();
});
const v = { canvas, ctx, scene, camera, controls, holder, spinner, labelEl, points: null, radius: 1 };
resizeViewer(v);
return v;
}
function clearViewer(v) {
if (v.points) {
v.scene.remove(v.points);
v.points.geometry.dispose();
v.points.material.dispose();
v.points = null;
}
}
// Drop gross outliers (e.g. GT "no-data" sentinels at 65.535 m) that would
// otherwise blow up the bounding box and push the real surface off-screen.
// Margins are generous, so clean clouds pass through untouched.
function robustClip(geometry) {
const pos = geometry.getAttribute('position');
const n = pos.count;
if (n < 100) return geometry;
// Percentile bounds per axis, estimated from a subsample for speed.
const STEP = Math.max(1, Math.floor(n / 20000));
const lo = [0, 0, 0], hi = [0, 0, 0];
for (let a = 0; a < 3; a++) {
const s = [];
for (let i = 0; i < n; i += STEP) s.push(pos.array[i * 3 + a]);
s.sort((p, q) => p - q);
const p01 = s[Math.floor(s.length * 0.01)];
const p99 = s[Math.floor(s.length * 0.99)];
const r = (p99 - p01) || 1e-6;
lo[a] = p01 - 2 * r; // keep a wide tail of legitimate edge points
hi[a] = p99 + 2 * r;
}
const col = geometry.getAttribute('color');
const keep = [];
for (let i = 0; i < n; i++) {
const x = pos.array[i * 3], y = pos.array[i * 3 + 1], z = pos.array[i * 3 + 2];
if (x >= lo[0] && x <= hi[0] && y >= lo[1] && y <= hi[1] && z >= lo[2] && z <= hi[2])
keep.push(i);
}
if (keep.length === n) return geometry; // nothing clipped
const np = new Float32Array(keep.length * 3);
const nc = col ? new Float32Array(keep.length * 3) : null;
for (let k = 0; k < keep.length; k++) {
const i = keep[k];
np[k * 3] = pos.array[i * 3]; np[k * 3 + 1] = pos.array[i * 3 + 1]; np[k * 3 + 2] = pos.array[i * 3 + 2];
if (nc) { nc[k * 3] = col.array[i * 3]; nc[k * 3 + 1] = col.array[i * 3 + 1]; nc[k * 3 + 2] = col.array[i * 3 + 2]; }
}
const g = new THREE.BufferGeometry();
g.setAttribute('position', new THREE.BufferAttribute(np, 3));
if (nc) g.setAttribute('color', new THREE.BufferAttribute(nc, 3));
return g;
}
// Place a (robust-clipped) cloud using a SHARED center + SHARED point size, so
// every panel stays co-registered in the sensor's metric frame at one scale.
function setGeometry(v, geometry, center, pointSize) {
clearViewer(v);
geometry.translate(-center.x, -center.y, -center.z);
const hasColor = !!geometry.getAttribute('color');
const material = new THREE.PointsMaterial({
size: pointSize,
vertexColors: hasColor,
color: hasColor ? 0xffffff : 0x4f9cff,
sizeAttenuation: true,
});
const points = new THREE.Points(geometry, material);
points.rotation.x = Math.PI; // most depth-backprojected clouds are y-down
v.scene.add(points);
v.points = points;
}
function loadPLY(file) {
return new Promise((res, rej) => loader.load(file, res, undefined, rej));
}
// Parse fov.json -> { vfov (deg), aspect = width/height }. Handles fov_x/fov_y
// (degrees) and/or a normalized 3x3 intrinsics matrix. Returns null if absent.
async function loadIntrinsics(url) {
try {
const r = await fetch(url);
if (!r.ok) return null;
const j = await r.json();
const K = Array.isArray(j.intrinsics) ? j.intrinsics : null;
let vfov = j.fov_y;
if (vfov == null && K) vfov = 2 * THREE.MathUtils.radToDeg(Math.atan(0.5 / K[1][1]));
if (vfov == null) return null;
let aspect;
if (j.fov_x != null) {
aspect = Math.tan(THREE.MathUtils.degToRad(j.fov_x) / 2) /
Math.tan(THREE.MathUtils.degToRad(vfov) / 2);
} else if (K) {
aspect = K[1][1] / K[0][0]; // fy/fx (both normalized) = width/height
} else {
aspect = 4 / 3;
}
return { vfov, aspect };
} catch (e) {
console.warn('No usable intrinsics at', url, e);
return null;
}
}
async function showScene(index) {
const scene = PC_SCENES[index];
const base = PC_BASE + scene.dir + '/';
// Update the RGB + sparse-sensor input thumbnails.
cmpImage.src = base + scene.image;
cmpSensor.src = base + scene.sensor;
// Fetch intrinsics in parallel with the point clouds.
const intrPromise = loadIntrinsics(base + (scene.fov || 'fov.json'));
// 1) Load + robust-clip every cloud up front (null on failure).
const geoms = await Promise.all(scene.panels.map(async (panel, i) => {
const v = viewers[i];
v.labelEl.textContent = panel.label;
v.spinner.style.display = 'flex';
v.spinner.textContent = 'Loading…';
try {
return robustClip(await loadPLY(base + panel.file));
} catch (e) {
console.error('Point cloud failed to load:', base + panel.file, e);
return null;
}
}));
// 2) Derive ONE shared center + radius from the reference cloud (Ours), so
// every panel is co-registered in the sensor frame and framed alike.
let refIdx = scene.panels.findIndex((p, i) => p.label === REF_LABEL && geoms[i]);
if (refIdx < 0) refIdx = geoms.findIndex((g, i) => g && scene.panels[i].kind !== 'input');
if (refIdx < 0) refIdx = geoms.findIndex(g => g);
const center = new THREE.Vector3();
let R = 1;
if (refIdx >= 0) {
const rg = geoms[refIdx];
rg.computeBoundingBox();
rg.boundingBox.getCenter(center);
rg.computeBoundingSphere();
R = rg.boundingSphere.radius || 1;
}
// 3) Place each cloud with the shared center + a single metric point size.
geoms.forEach((geom, i) => {
const v = viewers[i];
const panel = scene.panels[i];
if (!geom) {
clearViewer(v);
v.spinner.style.display = 'flex';
v.spinner.textContent = 'Not found';
return;
}
const sizeScale = panel.kind === 'input' ? SPARSE_SIZE : POINT_SIZE;
setGeometry(v, geom, center, R * sizeScale);
v.spinner.style.display = 'none';
});
// 4) Initial viewpoint. With intrinsics (fov.json) we reproduce the original
// image view: the virtual camera sits where the sensor was — the origin,
// which maps to world (-Cx, Cy, Cz) after our -center shift + x-flip — and
// looks down the optical axis (-z, up +y) with the sensor's vertical FOV
// and image aspect. Without intrinsics, fall back to a gentle framed orbit.
const intr = await intrPromise;
// Panel shape: default to the image aspect, but a scene may override it
// (e.g. KITTI's ~3.4:1 letterbox -> taller panel; the wide scene is cropped
// horizontally but the panels are much bigger). Keep the top input cards on
// the same aspect so row 1 stays aligned.
const displayAspect = scene.aspect ?? (intr ? intr.aspect : 4 / 3);
cmpImage.style.aspectRatio = String(displayAspect);
cmpSensor.style.aspectRatio = String(displayAspect);
if (intr) {
for (const v of viewers) {
v.holder.style.aspectRatio = String(displayAspect);
v.camera.fov = intr.vfov; // vertical FOV from intrinsics
v.camera.near = 0.03;
v.camera.far = Math.max(500, (center.z + R) * 8);
v.camera.up.set(0, 1, 0);
v.camera.position.set(-center.x, center.y, center.z);
v.controls.target.set(-center.x, center.y, 0);
resizeViewer(v); // applies new aspect + projection
v.controls.update();
}
} else {
const dist = R / Math.sin(THREE.MathUtils.degToRad(FOV / 2)) * DIST_PAD;
const px = dist * Math.sin(INIT_AZIM) * Math.cos(INIT_ELEV);
const py = dist * Math.sin(INIT_ELEV);
const pz = dist * Math.cos(INIT_AZIM) * Math.cos(INIT_ELEV);
for (const v of viewers) {
v.holder.style.aspectRatio = String(displayAspect);
v.camera.fov = FOV;
v.camera.near = dist / 100;
v.camera.far = dist * 100;
v.camera.up.set(0, 1, 0);
v.camera.position.set(px, py, pz);
v.controls.target.set(0, 0, 0);
resizeViewer(v);
v.controls.update();
}
}
renderAll();
}
function buildSelector() {
PC_SCENES.forEach((s, i) => {
const btn = document.createElement('button');
btn.className = 'comp-scene-btn' + (i === 0 ? ' active' : '');
btn.textContent = s.name;
btn.addEventListener('click', () => {
selector.querySelectorAll('.comp-scene-btn').forEach(b => b.classList.remove('active'));
btn.classList.add('active');
showScene(i);
});
selector.appendChild(btn);
});
}
// init — create the one shared WebGL context up front. If WebGL is
// unavailable (e.g. hardware acceleration disabled), degrade gracefully with
// a clear message instead of throwing and leaving an empty section.
try {