Feed-forward multi-view transformers such as VGGT recover the cameras and geometry of a video in one pass, but the cost of their cross-frame self-attention grows quadratically with the frame count, and their accuracy drops on long videos. Most remedies approximate attention over all frames under a fixed computational budget or select a few keys by their content. We hypothesize that the keys of a long video can be selected independently of video content, and our experiments on VGGT and π3 support it: a small number of uniformly sampled frames can serve as the keys of every frame. We present Eco3R, a training-free inference framework whose keys form a Pinned Key Set of 31 such frames. A pin pass stores their tokens in a Scene Cache, and the other frames read it in independent batches that reproduce joint inference with these keys exactly, at a constant cost per frame. An Online Stream updates the pinned frames during recording. Six middle global blocks align the frames and form a reusable VGGT module: the Alignment Score reads their attention to flag failures, and Alignment Refocusing reduces them by fine-tuning only these blocks (3.6 GPU hours). Eco3R (π3) matches π3's success rate on 1,000-frame ScanNet videos while running 13.8 times faster. With refocusing, Eco3R (VGGT) raises VGGT's success rate from 70% to 82% while running 8.5 times faster on 500 frames. Eco3R can reconstruct sequences of up to 4,000 frames on a single A100 GPU, while Eco3R (π3) streams 1,000 frames at 18.6 fps, achieving the lowest online absolute trajectory error.
Figures & tables
Figure 2: Three insights into cross-frame attention on long videos. (a) Median relative rotation error across 50 ScanNet scenes vs. the number of uniformly sampled keyframes, with all frames used as queries. (b) Increase of the median relative translation error at 500 frames when batches of 64 consecutive frames add 4, 8, or 16 of their own frames to 31 uniformly sampled keys. (c) Trajectory error of the full VGGT at 500 frames against the 90th percentile of the divergence of block 13 from uniform attention, one point per scene. AUROC, area under the receiver operating characteristic curve.
Figure 3: Overview of Eco3R. A pin pass processes the Pinned Key Set of uniformly sampled frames (dark blue) and stores the inputs of every global block in the Scene Cache. In attach passes, the other frames (light blue) read the keys KP and values VP only from this Scene Cache. Alignment Refocusing fine-tunes blocks 10–15, while the Alignment Score α is computed from block 13.
Figure 4: The alignment band. (a) Divergence from uniform attention per global block. (b) Rank correlation between the attention share of two distant frames and their co-visibility, with the interquartile range of VGGT in gray.
Figure 5: Accuracy and cost on ScanNet. (a) Success rate vs. speed (fps) at 1,000 frames. The dotted line denotes the Pareto front of competing methods, with the region below it shaded and the corner shown in an enlarged inset. GTH is Good Token Hunting. (b) Median time per frame and (c) peak inference memory against video length.
500 frames
1,000 frames
Method
Success ↑
ATE ↓
RPE ↓
Acc. ↓
Time ↓
Mem. ↓
Success ↑
ATE ↓
RPE ↓
Acc. ↓
Time ↓
Mem. ↓
VGGT (full model) [ 51 ]
70
0.111
0.589
1.97
196
23.1
70
0.130
0.581
2.26
772
44.2
π3 (full model) [ 56 ]
94
0.054
0.343
0.94
138
24.4
94
0.058
0.265
0.90
557
43.4
Offline
FastVGGT [ 39 ]
74
0.098
0.598
1.95
68.8
32.3
72
0.100
0.521
2.00
196
62.5
FlashVGGT [ 57 ] †
76
0.121
0.563
1.88
19.7
38.0
76
0.118
0.426
1.80
50.6
70.1
Table 1: Long ScanNet videos on one A100 GPU. Units are meters for absolute trajectory error (ATE), degrees for relative rotation error (RPE), centimeters for point accuracy (Acc.), seconds per scene for runtime, and GB for peak inference memory (maximum across scenes). Shaded rows are Eco3R (ours), while bold values denote the best results in each group. “Failed” indicates a success rate of at most 20%, with the rate shown in parentheses. † Trains or fine-tunes network weights. ‡ Full pipeline (inference 74 s and 151 s).
2,000 frames
4,000 frames
Method
Succ. ↑
ATE ↓
ms/fr. ↓
Mem. ↓
Succ. ↑
ATE ↓
ms/fr. ↓
Mem. ↓
VGGT [ 51 ]
42
0.122
846
> 80
42
0.122
846
> 80
π3 [ 56 ]
60
0.050
641
> 80
60
0.050
641
> 80
Offline
FlashVGGTStream [ 57 ] †
failed (14)
—
—
—
failed (14)
—
—
—
VGGT-Long [ 14 ] ‡
56
0.169
566
20.4
56
0.171
565
20.4
Table 2: Thousands of frames on ScanNet. Units and notation follow Tab. 1 , with time per frame in ms. Scenes that exceed 80 GB of memory count as failures, and ATE and time are medians over the remaining scenes.
Succ. ↑
TFLOPs/frame ↓
FPS ↑
Latency (s) ↓
Energy ↓
Method
(%)
500
1,000
500
1,000
Median
First
(J/frame)
VGGT [ 51 ]
70
55.6
108.9
2.6
1.3
822
872
223.4
π3 [ 56 ]
94
42.2
82.1
3.6
1.8
607
656
167.0
Offline
FastVGGT [ 39 ]
72
13.7
25.1
7.3
5.1
246
296
51.7
FlashVGGT [ 57 ] †
76
4.1
5.9
25.3
19.8
101
150
14.7
Table 3: Efficiency on ScanNet on one A100 GPU. TFLOPs count a multiply-add as two floating-point operations. Latency is the median and first-frame delay when 1,000 frames arrive at 10 per second. Success rate and GPU energy are measured at 1,000 frames. OOM marks out of memory, and other notation follows Tab. 1 .
Figure 6: Two baselines and Eco3R on two long ScanNet videos , scene0648_00 (top) and scene0477_00 (bottom), each in one oblique view. Icons mark which frames (columns) each frame (rows) attends to: a pruned cache, overlapping chunks, or the same pinned frames.
7-Scenes
NRGBD
TUM RGB-D
Replica
Method
Acc.
ATE
Acc.
ATE
Acc.
ATE
Acc.
ATE
VGGT [ 51 ]
1.75
0.124
1.91
0.052
1.43
0.076
4.70
0.182
π3 [ 56 ]
1.07
0.088
1.27
0.030
0.99
0.073
1.08
0.030
Offline
FastVGGT [ 39 ]
1.62
0.127
1.78
0.045
1.51
0.079
3.38
0.098
VGGT-Long [ 14 ]
1.77
0.159
1.82
0.075
1.93
0.127
1.82
0.078
Table 4: Four other datasets. Units and notation follow Tab. 1 . ∗ Stops on 1 of 18 sequences of 7-Scenes and 2 of 12 sequences of TUM RGB-D, and the medians exclude these sequences.
Variant
Succ. ↑
ATE ↓
RPE ↓
Acc. ↓
Comp. ↓
AbsRel ↓
TFLOPs ↓
FPS ↑
Eco3R (VGGT) †
82
0.090
0.528
1.48
1.46
0.017
5.58
21.6
w/o Alignment Refocusing
− 10
+ 29.7
+ 8.1
+ 35.0
+ 20.9
+ 29.0
1.00
1.00
w/o Pinned Key Set
− 4
+ 19.8
+ 7.8
+ 11.9
+ 3.5
+ 3.7
9.96
0.13
w/o both (VGGT)
− 12
+ 23.8
+ 11.4
+ 32.8
+ 22.0
+ 28.8
9.96
0.12
16 pinned frames
− 4
+ 13.9
+ 10.1
+ 9.9
+ 3.5
+ 6.2
0.71
1.30
63 pinned frames
0
+ 0.8
− 3.4
+ 0.3
− 1.3
− 3.4
1.62
0.68
Table 5: Ablations of Eco3R on ScanNet at 500 frames. Shaded rows give absolute values in the units and notation of Tabs. 1 and 3 (Comp. in cm). The rows below each shaded row give the changes in success rate (points) and median errors (%), as well as the ratios of TFLOPs and FPS. Indented rows compare with the Online Stream row. Gray marks changes below 5% that are not significant ( p≥0.05 ).
Figure S1 : Online Stream. Three consecutive passes for κ=16 , Mmax=6 , and τ=2 . A new candidate key arrives every κ frames, and each pass runs a new pin pass over at most Mmax candidates sampled uniformly from the past, so the pinned keys shift as the video grows. Each pass attaches and emits the κ frames up to the candidate that arrived τ passes earlier, and the pinned frames that earlier passes emitted set its scale. A frame waits at most (τ+1)κ frames before it is emitted.
Method
Succ. ↑
ATE ↓
RPE-r ↓
RPE-t ↓
AbsRel ↓
Acc. ↓
Comp. ↓
Time ↓
Mem. ↓
500 frames
VGGT [ 51 ]
70
0.111
0.589
2.59
0.0218
1.97
1.78
196
23.1
π3 [ 56 ]
94
0.054
0.343
1.47
0.0109
0.94
1.15
138
24.4
DA3 [ 26 ]
64
0.137
1.004
3.62
0.0269
2.61
2.15
66.1
—
Offline
FastVGGT [ 39 ]
74
0.098
0.598
2.58
0.0222
1.95
1.73
68.8
32.3
Table S1 : ScanNet at 500 and 1,000 frames , extending Tab. 1 . Succ. is the success rate in %, and the absolute relative depth error (AbsRel) has no unit. Other units are meters for the absolute trajectory error (ATE), degrees and centimeters for the relative rotation and translation errors (RPE-r and RPE-t), centimeters for accuracy (Acc.) and completeness (Comp.), seconds for the median time per scene, and GB for the largest peak memory of inference over the scenes ( Sec. S2 ). The first three rows at each length are the full models, and other notation follows Tab. 1 , except that failed marks a success rate below 10%. Dashes mark the errors of failed rows and the memory that we did not measure. ‡ Time of the full pipeline, including 74 s and 151 s of inference at 500 and 1,000 frames. ∥ Run with unified memory in the 84% of scenes that exceed 80 GB at 1,000 frames, so we omit its time and memory. Scenes on which FrameVGGT (11 and 24) or VGGT-SLAM (1 and 6) stops early count as failures ( Sec. S2 ).
500 frames
1,000 frames
Method
Succ. ↑
ATE ↓
RPE ↓
Acc. ↓
Time ↓
Mem. ↓
Succ. ↑
ATE ↓
RPE ↓
Acc. ↓
Time ↓
Mem. ↓
VGGT [ 51 ]
70
0.111
0.589
1.97
196
23.1
70
0.130
0.581
2.26
772
44.2
π3 [ 56 ]
94
0.054
0.343
0.94
138
24.4
94
0.058
0.265
0.90
557
43.4
VGGT backbone
Wang et al. [ 49 ] ∥
62
0.113
0.673
2.43
85.4
41.2
62
0.132
0.678
2.53
—
—
SAF3R [ 13 ] ∥
74
0.110
0.543
1.94
56.7
39.9
72
0.114
0.469
2.07
—
—
Table S2 : Sparse attention and key selection on ScanNet. Succ. is in %, RPE is the relative rotation error (RPE-r in Tab. S1 ), and other units are as in Tab. S1 . The first two rows are the full models, shaded rows are Eco3R, and bold marks the best value in each column of each backbone group. ∥ Run with unified memory in the 78% of scenes that exceed 80 GB at 1,000 frames, so we omit its time and memory.
Method
Succ.
ATE
RPE-r
RPE-t
AbsRel
Acc.
Time
496 frames
VGGT [ 51 ]
68
0.111
0.590
2.57
0.0219
1.99
92.2
FastVGGT [ 39 ]
74
0.099
0.597
2.58
0.0223
1.96
24.1
LiteVGGT [ 41 ] †
68
0.127
0.677
3.09
0.0235
2.31
13.0
Eco3R (VGGT) †
82
0.094
0.537
1.86
0.0165
1.48
11.8
1,000 frames
Table S3 : LiteVGGT on one H100 GPU. All methods use 496 in place of 500 frames. Succ. is in %, AbsRel has no unit, and other units are as in Tab. S1 . Shaded rows are Eco3R, and bold marks the best value at each length.
Method
Frames
Succ.
ATE
RPE
Acc.
ms/fr.
Mem.
Offline
Eco3R (VGGT) †
2,000
82
0.101
0.431
1.54
46.7
10.0
4,000
82
0.098
0.441
1.54
46.7
10.0
Eco3R ( π3 )
2,000
92
0.060
0.270
1.03
40.6
9.9
4,000
90
0.060
0.275
1.03
40.0
9.9
Online
Table S4 : Eco3R at thousands of frames on ScanNet. Units are as in Tab. S1 . RPE is the relative rotation error, ms/fr. is the median time per frame, and Mem. is the largest peak memory of inference over the scenes.
500 frames
1,000 frames
Method
Succ. prot.
Succ. std.
ATE std.
Succ. prot.
Succ. std.
ATE std.
VGGT [ 51 ]
70
96
0.079
70
88
0.084
π3 [ 56 ]
94
100
0.039
94
100
0.038
Offline
FastVGGT [ 39 ]
74
98
0.074
72
96
0.079
FlashVGGT [ 57 ] †
76
94
0.080
76
92
0.081
Table S5 : Protocol and standard ATE on ScanNet. Succ. prot. and Succ. std. are the shares of scenes with an ATE below 0.2 m (%) under the protocol of FastVGGT and under the standard ATE, respectively. ATE std. is the median standard ATE in meters. The first two rows are the full models, shaded rows are Eco3R, and bold marks the best value in each column of the Offline and Online groups.
Method
500 frames
1,000 frames
2,000 frames
4,000 frames
VGGT (full model) [ 51 ]
70 [56, 81]
70 [56, 81]
42 [29, 56]
42 [29, 56]
π3 (full model) [ 56 ]
94 [84, 98]
94 [84, 98]
60 [46, 72]
60 [46, 72]
Offline
FastVGGT [ 39 ]
74 [60, 84]
72 [58, 83]
—
—
FlashVGGT [ 57 ] †
76 [63, 86]
76 [63, 86]
—
—
FlashVGGTStream [ 57 ] †
26 [16, 40]
18 [10, 31]
14 [7, 26]
14 [7, 26]
Table S6 : Success rates with 95% confidence intervals on ScanNet (%), for the rows of Tabs. 1 and 2 , with Wilson intervals over the 50 scenes in brackets. Rows marked failed in Tabs. 1 and 2 keep their rates here, OOM marks out of memory, and dashes mark methods that Tab. 2 omits.
7-Scenes
NRGBD
TUM RGB-D
Replica
Method
Acc.
Comp.
ATE
RPE
Time
Acc.
Comp.
ATE
RPE
Time
Acc.
Comp.
ATE
RPE
Time
Acc.
Comp.
ATE
RPE
Time
VGGT [ 51 ]
1.75
2.04
0.124
0.41
195.3
1.91
1.30
0.052
0.22
136.5
1.43
1.62
0.076
0.64
325.0
4.70
2.51
0.182
0.69
437.8
π3 [ 56 ]
1.07
1.68
0.088
0.34
138.8
1.27
0.76
0.030
0.07
205.7
0.99
1.40
0.073
0.49
232.2
1.08
0.76
0.030
0.10
313.4
Offline
FastVGGT [ 39 ]
1.62
2.04
0.127
0.41
56.5
1.78
1.30
0.045
0.20
106.0
1.51
1.44
0.079
0.63
130.3
3.38
2.18
0.098
0.42
134.6
VGGT-Long [ 14 ]
1.77
2.11
0.159
0.36
274.8
1.82
1.15
0.075
0.09
266.9
1.93
1.57
0.127
0.48
358.3
1.82
1.40
0.078
0.12
344.6
Table S7 : Other datasets. Medians over the sequences of each dataset: point accuracy and completeness in centimeters, ATE in meters (protocol of FastVGGT), relative rotation error (RPE) in degrees, and time per sequence in seconds. The first two rows are the full models, shaded rows are Eco3R, bold marks the best value in each column of the Offline and Online groups, and failed marks a success rate below 10% on that dataset. ∗ Stops on 1 of 18 sequences of 7-Scenes and 2 of 12 sequences of TUM RGB-D, and the medians exclude these sequences.
Latency ↓
Method
Rate ↑
Median
Max.
First ↓
Energy ↓
CUT3R [ 53 ] †
failed
—
—
—
—
TTT3R [ 6 ]
failed
—
—
—
—
RetrieveVGGT [ 66 ]
1.2
346.4
755.8
0.26
106.4
InfiniteVGGT [ 62 ]
3.8
72.8
162.5
0.28
61.7
LASER [ 16 ]
11.3
3.9
5.7
5.30
20.1
Table S8 : Streaming 1,000 frames at 10 frames per second on one A100 GPU. Rate is the number of frames processed per second without waiting, measured on the streamed videos. Median, Max., and First give the median, maximum, and first-frame latency in seconds, and Energy is the GPU energy per frame in joules. Shaded rows are Eco3R, bold marks the best value in each column, and failed marks a success rate below 10% at 1,000 frames.
Figure S2 : Latency of every frame when 1,000 frames arrive at 10 per second on one A100 GPU, median over the streamed videos that reach each frame (ten, or five for InfiniteVGGT).
Method
500 frames
1,000 frames
VGGT
110.8
223.4
π3
82.9
167.0
Offline
FastVGGT
34.2
51.7
FlashVGGT †
11.6
14.7
VGG-T 3 †
—
16.1
Table S9 : GPU energy per frame in joules on one A100 GPU, median over scenes. The first two rows are the full models, bold marks the lowest energy in each column of the Offline and Online groups, failed marks a success rate below 10%, and dashes mark settings without an energy measurement. ∗ Most of its runtime is on the CPU, which this GPU measurement excludes.
Frames serving as keys
Backbone
Length
16
31
63
125
All
VGGT
500
70 / 0.642
72 / 0.571
74 / 0.549
72 / 0.549
70 / 0.589
VGGT
1,000
72 / 0.510
72 / 0.501
72 / 0.471
72 / 0.454
70 / 0.581
π3
500
90 / 0.383
94 / 0.356
94 / 0.347
94 / 0.345
94 / 0.343
π3
1,000
92 / 0.302
94 / 0.281
94 / 0.271
94 / 0.267
94 / 0.265
Table S10 : Number of pinned frames on ScanNet. Success rate (%) and median relative rotation error (°) of the frozen VGGT and of π3 when the keys of every global block come from the given number of uniformly sampled frames or from all frames, as in Fig. 2 a. For the frozen VGGT, the keys are a fixed share of each video. Videos that reach the frame budget get the stated numbers, with 62 in place of 63, and shorter videos get proportionally fewer keys.
Variant
Succ. ↑
ATE ↓
RPE ↓
Acc. ↓
Comp. ↓
AbsRel ↓
TFLOPs ↓
FPS ↑
Eco3R (VGGT) †
82
0.090
0.528
1.48
1.46
0.0170
5.58
21.6
Components
w/o Alignment Refocusing
72
0.117
0.571
2.00
1.76
0.0219
5.58
21.5
w/o Pinned Key Set
78
0.108
0.569
1.66
1.51
0.0176
55.59
2.7
w/o both (VGGT)
70
0.111
0.589
1.97
1.78
0.0218
55.59
2.6
Pinned Key Set
Table S11 : Ablations of Eco3R on ScanNet at 500 frames in absolute values, in the units of Tabs. 1 and 3 (Comp. in cm). Gray values do not differ significantly ( p≥0.05 ) from their reference, regardless of the size of the change. Most rows compare with the shaded row above them, and the indented rows compare with the Online Stream row. The nearby-keys and camera-head rows compare with their shaded configuration rerun on the same 24 GB workstation GPU. ∗ Compared with a band trained with the full loss on the same 482 scans. Dashes mark costs that we do not report.
Layout
Succ.
ATE
RPE-r
RPE-t
Acc.
Eco3R (VGGT) †
Uniformly sampled
82
0.090
0.528
1.92
1.48
Random, draw 1
80
0.101
0.563
1.90
1.58
Random, draw 2
80
0.107
0.541
1.99
1.52
Random, draw 3
80
0.114
0.578
2.00
1.60
Random, mean of draws
—
0.108
0.574
2.00
1.57
Table S12 : Layout of the pinned frames on ScanNet at 500 frames. Succ. is the success rate (%), and the errors are medians over the scenes in the units of Tab. S1 . Gray values do not differ significantly from those of uniformly sampled frames ( p≥0.05 ). The mean row averages each error over the three random draws of each scene. All rows ran on the same 24 GB workstation GPU.
Variant
Succ.
ATE
RPE-r
RPE-t
Acc.
Mean
One block reverted
Block 10
82
98
67
90
93
87
Block 11
82
89
94
87
93
91
Block 12
80
117
66
71
86
85
Block 13
82
102
81
97
102
96
Block 14
78
95
93
90
95
93
Table S13 : Where Alignment Refocusing acts on ScanNet at 500 frames. Each variant reverts some tuned weights of the band to those of the frozen VGGT at inference. Succ. is the success rate (%). The error columns divide the change of each median error against Eco3R (VGGT) without refocusing by that of Alignment Refocusing, giving the share of the gain that the variant retains (%). Mean averages the four shares. Eco3R (VGGT) succeeds in 72% of the scenes without refocusing and in 82% with it. All runs used the same 24 GB workstation GPU.
7-Scenes
NRGBD
TUM RGB-D
Replica
Method
Acc.
Comp.
ATE
RPE
Acc.
Comp.
ATE
RPE
Acc.
Comp.
ATE
RPE
Acc.
Comp.
ATE
RPE
VGGT [ 51 ]
1.75
2.04
0.124
0.41
1.91
1.30
0.052
0.22
1.43
1.62
0.076
0.64
4.70
2.51
0.182
0.69
Eco3R (VGGT) w/o Alignment Refocusing
1.74
2.03
0.124
0.39
1.34
1.13
0.030
0.18
1.24
1.69
0.062
0.65
2.04
1.62
0.075
0.31
Eco3R (VGGT) †
1.18
2.04
0.100
0.38
1.77
1.32
0.055
0.19
1.10
1.37
0.091
0.62
3.90
2.64
0.254
0.38
Table S14 : Pinned Key Set and Alignment Refocusing on other datasets. Medians over the sequences of each dataset, in the units of Tab. S7 . Bold marks the best value in each column.
Variant
Succ. ↑
ATE ↓
RPE ↓
Acc. ↓
ms/fr. ↓
Mem. ↓
500 frames
Eco3R (VGGT- Ω )
76
0.112
0.410
1.95
23.5
10.9
w/o Pinned Key Set
74
0.107
0.417
1.97
128
30.5
Online
74
0.116
0.413
2.16
49.4
9.4
1,000 frames
Eco3R (VGGT- Ω )
76
0.106
0.322
1.98
23.7
10.9
Table S15 : Eco3R on VGGT- Ω on ScanNet. The shaded rows are Eco3R (VGGT- Ω ) offline, as in Tab. 5 . The row without the Pinned Key Set is the full VGGT- Ω , which we run up to 1,000 frames, and the Online row is the Online Stream on this backbone. Time is per frame in ms, other units and notation follow Tabs. 1 and 2 , and the runs at 2,000 and 4,000 frames skip the point clouds.
7-Scenes
NRGBD
TUM RGB-D
Replica
Method
Acc.
ATE
Acc.
ATE
Acc.
ATE
Acc.
ATE
Eco3R (VGGT- Ω )
1.96
0.136
1.47
0.022
1.03
0.051
1.24
0.014
w/o Pinned Key Set
2.02
0.139
1.65
0.023
0.99
0.052
1.38
0.028
Online
2.04
0.136
1.65
0.022
1.43
0.058
1.30
0.016
Table S16 : Eco3R on VGGT- Ω on the other datasets, with point accuracy (Acc., cm) and ATE (m) as in Tab. 4 and the rows of Tab. S15 .
Figure S3 : Three ways to run a multi-view transformer on a long video, and Eco3R on the same video. All panels show the same ScanNet video of a bedroom (scene0477_00) from above, without the ceiling. Lines show the true camera path in the first panel and the predicted paths in the others. Icons mark which frames (columns) each frame (rows) attends to, titles give the number of frames, and the label under each reconstruction gives its ATE. Top row: with all frames as keys, every prediction lies in one coordinate frame, but the computation grows quadratically with the number of frames. Chunk and stitch bounds the cost, but errors accumulate along the chain of windows. When the camera returns to a wall, the last frames duplicate the wall (red ellipses). Pruned streaming reconstructs the video with bounded GPU memory as it arrives, but later frames align to the cached tokens of earlier frames, including incorrectly positioned ones. The wall that frames 100–199 incorrectly positioned remains uncorrected (red ellipse), and the room is reconstructed as rotated copies. Bottom row: Eco3R (VGGT) at 1,000 frames, Eco3R ( π3 ) offline and online at 1,000 frames, and Eco3R ( π3 ) offline at 4,000 frames.
Figure S4 : More videos. Top two rows: scene0000_00 in top views, with the ground truth, the full VGGT, LASER, and RetrieveVGGT in the first row and the four runs of Eco3R in Fig. S3 in the second. Lines show the camera paths, and icons show which frames each frame attends to, as in Fig. S3 . Bottom row: the top video of Fig. 6 in its oblique view, with two input frames, the RGB-D ground truth, and the same four runs of Eco3R. Each title gives the number of frames, and the label under each reconstruction gives its ATE.
Eco3R, Scene Cache of 31 frames
Scene
(VGGT) w/o Alignment Refocusing
(VGGT) †
( π3 )
Chess
3.9 / 1.83
3.6 / 1.75
3.7 / 1.83
Fire
2.9 / 1.96
2.8 / 1.79
3.5 / 2.11
Heads
1.6 / 1.92
1.5 / 2.09
1.8 / 2.07
Office
8.9 / 3.64
6.1 / 3.18
7.9 / 4.72
Pumpkin
8.1 / 3.22
6.9 / 2.91
5.7 / 2.84
Table S17 : Relocalization on 7-Scenes. Median translation (cm) and rotation (°) errors of the test images in each scene, and their mean over the seven scenes.
Eco3R ( π3 )
Sequence
Frames
Online
Recent keys
Offline
LongStream [ 8 ] †
00
4,541
179.7
177.2
189.6
89.3
01
1,101
405.9
256.7
282.9
63.6
02
4,661
301.8
265.8
301.5
236.9
03
801
41.6
21.6
6.7
5.2
04
271
6.9
6.9
2.3
2.5
Table S18 : KITTI odometry. ATE (m) after alignment by a similarity transform alone. The columns under Eco3R ( π3 ) give the Online Stream with κ=16 and Mmax=32 , the same stream with keys from the 32 most recent candidates, and the offline Pinned Key Set with 31 keys.
University of Chinese Academy of Sciences, Beijing, China · the University of Maryland, College Park, MD, USA · Beijing University of Posts and Telecommunications, Beijing, China