DAGE: Dual-Stream Architecture for Efficient and Fine-Grained Geometry Estimation

Tuan Duc Ngo; Jiahui Huang; Seoung Wug Oh; Kevin Blackburn-Matzen; Evangelos Kalogerakis; Chuang Gan; Joon-Young Lee

cs.CVMar 4, 2026

DAGE: Dual-Stream Architecture for Efficient and Fine-Grained Geometry Estimation

Authors: Tuan Duc Ngo, Jiahui Huang, Seoung Wug Oh, Kevin Blackburn-Matzen, Evangelos Kalogerakis, Chuang Gan, Joon-Young Lee

Organizations: UMass Amherst · Adobe Research · TU Crete

Abstract

Estimating accurate, view-consistent geometry and camera poses from uncalibrated multi-view/video inputs remains challenging - especially at high spatial resolutions and over long sequences. We present DAGE, a dual-stream transformer whose main novelty is to disentangle global coherence from fine detail. A low-resolution stream operates on aggressively downsampled frames with alternating frame/global attention to build a view-consistent representation and estimate cameras efficiently, while a high-resolution stream processes the original images per-frame to preserve sharp boundaries and small structures. A lightweight adapter fuses these streams via cross-attention, injecting global context without disturbing the pretrained single-frame pathway. This design scales resolution and clip length independently, supports inputs up to 2K, and maintains practical inference cost. DAGE delivers sharp depth/pointmaps, strong cross-view consistency, and accurate poses, establishing new state-of-the-art results for video geometry estimation and multi-view reconstruction.

Figures & tables

Figure 2: Overview of DAGE. Given a set of unposed RGB images, the model predicts per-frame pointmaps and camera poses, plus a scene-wise metric scale. The architecture has two parallel streams: (i) a low-resolution stream (lower part) that processes downsampled inputs to aggregate global context and regress poses/scene scale; and (ii) a high-resolution stream (upper part) that processes frames independently at native resolution to preserve fine detail. A lightweight Adapter fuses LR and HR tokens before the dense geometry head.

Figure 3: The Global transformer (left) operates on low-resolution inputs with alternating global and frame-wise attention; during training, feature distillation compensates for aggressive downsampling. The Adapter (right) stacks cross and self-attention blocks to fuse multi-view–consistent LR tokens into the HR stream.

Method	MV	HR	PO	GMU [ 34 ]		Monkaa [ 64 ]		Sintel [ 9 ]		ScanNet [ 20 ]		KITTI [ 33 ]		UrbanSyn [ 37 ]		Unreal4K [ 92 ]		Diode [ 93 ]		Rank $\downarrow$
				( $960\times 512$ )		( $960\times 512$ )		( $896\times 448$ )		( $640\times 512$ )		( $768\times 384$ )		( $2048\times 1024$ )		( $1920\times 1080$ )		( $1024\times 768$ )
				$\mathrm{Rel}^{p}\!\downarrow$	$\delta^{p}\!\uparrow$	$\mathrm{Rel}^{p}\!\downarrow$	$\delta^{p}\!\uparrow$	$\mathrm{Rel}^{p}\!\downarrow$	$\delta^{p}\!\uparrow$	$\mathrm{Rel}^{p}\!\downarrow$	$\delta^{p}\!\uparrow$	$\mathrm{Rel}^{p}\!\downarrow$	$\delta^{p}\!\uparrow$	$\mathrm{Rel}^{p}\!\downarrow$	$\delta^{p}\!\uparrow$	$\mathrm{Rel}^{p}\!\downarrow$	$\delta^{p}\!\uparrow$	$\mathrm{Rel}^{p}\!\downarrow$	$\delta^{p}\!\uparrow$
DepthPro [ 8 ]		✓		9.5	93.9	25.1	58.4	40.8	44.7	9.3	94.9	10.0	94.9	48.9	40.1	74.7	12.0	32.4	59.2	7.9
MoGe [ 99 ]		✓		20.3	71.2	22.9	61.3	29.4	59.8	13.4	88.0	8.0	95.8	14.9	87.0	38.3	51.5	31.8	52.9	7.4
MoGe2 [ 100 ]		✓		19.6	72.4	25.0	57.0	29.8	58.4	12.4	89.4	9.0	97.2	13.4	90.0	32.9	59.1	31.0	54.2	6.8
MoGe2 [ 100 ] †	$\triangle$	✓		7.1	94.6	21.4	67.6	28.2	62.8	7.8	97.5	10.5	98.4	7.2	97.1	12.6	86.7	15.8	84.1	4.1

Table 1: Video pointmap evaluation . Results are aligned with the ground truth by optimizing a shared scale and shift factor across the entire video. “MV/HR/PO” indicate multi-view support, high-resolution input support, and whether the method predicts camera poses. We mark best and second-best .

Figure 4: Visual comparison of video depth on in-the-wild scenes. We convert the depth map to a disparity map for better visualization, and zoom-in ( red bounding boxes ) to emphasize details. DAGE preserves sharp boundaries and fine-grained detail—especially for thin structures and small or distant objects, outperforming a diffusion-based baseline [ 109 ] .

Figure 5: Visual comparison of 3D reconstruction on in-the-wild scenes. Compared to VGGT [ 96 ] and Pi3 [ 103 ] , DAGE achieves comparable multi-view consistency while preserving markedly finer detail ( green boxes ).

Method	Monkaa [ 64 ]		Sintel [ 9 ]		UrbanSyn [ 37 ]		Unreal4K [ 92 ]
Method	F1 $\uparrow$	$C_{\mathrm{PDBE}}\downarrow$	F1 $\uparrow$	$C_{\mathrm{PDBE}}\downarrow$	F1 $\uparrow$	$C_{\mathrm{PDBE}}\downarrow$	F1 $\uparrow$	$C_{\mathrm{PDBE}}\downarrow$
DepthPro [ 8 ]	0.19	21.3	0.41	17.0	0.14	12.5	0.07	116.4
MoGe2 [ 100 ]	0.27	11.6	0.27	10.1	0.09	19.1	0.10	35.2
GeoCrafter [ 109 ]	0.19	8.86	0.28	8.1	0.08 $\star$	33.2 $\star$	0.06 $\star$	41.4 $\star$
CUT3R [ 98 ]	0.08	20.3	0.11	16.5	0.01	44.0	0.01	63.1
VGGT [ 96 ]	0.14	11.1	0.20	9.6	0.02 $\star$	42.0 $\star$	0.03 $\star$	38.1 $\star$

Table 2: Sharpness depth evaluation .

Method	Setting	7-Scenes [ 82 ]			NRGBD [ 2 ]
Method	Setting	Acc. $\downarrow$	Comp. $\downarrow$	NC $\uparrow$	Acc. $\downarrow$	Comp. $\downarrow$	NC $\uparrow$
Fast3R [ 110 ]	sparse	0.065	0.089	0.759	0.091	0.104	0.877
CUT3R [ 98 ]		0.049	0.051	0.805	0.041	0.031	0.968
FLARE [ 120 ]		0.057	0.107	0.780	0.024	0.025	0.988
VGGT [ 96 ]		0.025	0.033	0.845	0.029	0.038	0.981
Pi3 [ 103 ]		0.029	0.049	0.841	0.015	0.014	0.992

Table 3: Multi-view reconstruction evaluation. We report the median values on 3 settings, including sparse , dense , and metric .

Method	Sintel [ 9 ]			TUM-dynamics [ 85 ]			ScanNet [ 20 ]
Method	ATE $\downarrow$	RPE T $\downarrow$	RPE R $\downarrow$	ATE $\downarrow$	RPE T $\downarrow$	RPE R $\downarrow$	ATE $\downarrow$	RPE T $\downarrow$	RPE R $\downarrow$
Fast3R [ 110 ]	0.371	0.298	13.75	0.090	0.101	1.425	0.155	0.123	3.491
CUT3R [ 98 ]	0.217	0.070	0.636	0.047	0.015	0.451	0.094	0.022	0.629
FLARE [ 120 ]	0.207	0.090	3.015	0.026	0.013	0.475	0.064	0.023	0.971
VGGT [ 96 ]	0.167	0.062	0.491	0.012	0.010	0.311	0.035	0.015	0.382
Pi3 [ 103 ]	0.074	0.040	0.282	0.014	0.009	0.312	0.031	0.013	0.347

Table 4: Camera pose evaluation

Resolution	MoGe2 †		GeoCrafter		CUT3R		VGGT		Pi3		DAGE
Resolution	FPS	Mem.	FPS	Mem.	FPS	Mem.	FPS	Mem.	FPS	Mem.	FPS	Mem.
540 $\times$ 360	79.4	8.1	3.1	17.3	27.2	16.5	30.1	17.3	36.3	17.2	65.4	10.1
960 $\times$ 512	30.0	15.3	1.7	24.1	20.3	19.0	2.1	26.9	3.1	23.1	28.9	18.3
2048 $\times$ 1024	6.1	22.1	OOM		4.5	33.2	OOM		0.2	66.7	5.6	27.9

Table 5: Throughput comparison. FPS

\uparrow

/ GPU memory

\downarrow

(GB) measured on 100-frame clips per resolution.

Table 6: Ablation studies (a) different adapter design, (b) effect of each component on the depth sharpness

Reso.	Sintel [ 9 ] (pose)			Sintel [ 9 ] (depth)			NRGBD [ 2 ]		FPS $\uparrow$
Reso.	ATE $\downarrow$	RPE T $\downarrow$	RPE R $\downarrow$	Rel ${}^{p}\!\downarrow$	$\delta^{p}\!\uparrow$	F1 $\uparrow$	Acc. $\downarrow$	Comp. $\downarrow$	FPS $\uparrow$
252px (default)	0.132	0.051	0.406	21.5	75.5	0.34	0.018	0.016	65.4
252px (no distill)	0.111	0.057	0.584	22.9	73.0	0.35	0.019	0.017	65.4
336px	0.130	0.042	0.307	20.4	76.9	0.35	0.018	0.016	46.6
518px	0.117	0.039	0.258	19.7	77.7	0.36	0.016	0.016	22.5

Table 7: Ablation study on the LR stream resolution

Figure 6: Comparison on disparity quality on Pi3, our variant without monocular prior, and our full model.

Dataset Name	Scene type	Metric	Real	Dynamic
ARKitScenes [ 3 ]	Indoor	Yes	Real	Static
ScanNet [ 20 ]	Indoor	Yes	Real	Static
ScanNet++ [ 116 ]	Indoor	Yes	Real	Static
TartanAir [ 102 ]	Mixed	Yes	Synthetic	Dynamic
Waymo [ 87 ]	Outdoor	Yes	Real	Dynamic
BlendedMVS [ 115 ]	Mixed	No	Synthetic	Static

Table 8: Training datasets .

Method	GMU [ 34 ]		Monkaa [ 64 ]		Sintel [ 9 ]		ScanNet [ 20 ]		KITTI [ 33 ]		UrbanSyn [ 37 ]		Unreal4K [ 92 ]		Diode [ 93 ]		Rank $\downarrow$
Method	$\mathrm{Rel}^{p}\!\downarrow$	$\delta^{p}\!\uparrow$	$\mathrm{Rel}^{p}\!\downarrow$	$\delta^{p}\!\uparrow$	$\mathrm{Rel}^{p}\!\downarrow$	$\delta^{p}\!\uparrow$	$\mathrm{Rel}^{p}\!\downarrow$	$\delta^{p}\!\uparrow$	$\mathrm{Rel}^{p}\!\downarrow$	$\delta^{p}\!\uparrow$	$\mathrm{Rel}^{p}\!\downarrow$	$\delta^{p}\!\uparrow$	$\mathrm{Rel}^{p}\!\downarrow$	$\delta^{p}\!\uparrow$	$\mathrm{Rel}^{p}\!\downarrow$	$\delta^{p}\!\uparrow$	Rank $\downarrow$
DepthPro [ 8 ]	10.5	92.7	27.9	51.2	55.0	37.5	9.3	95.0	11.7	93.6	22.5	61.1	96.1	1.2	30.3	58.1	7.6
MoGe [ 99 ]	21.4	69.0	27.7	58.3	29.5	59.8	13.4	88.2	8.6	95.6	13.4	89.9	34.0	55.7	30.3	53.4	6.9
MoGe2 [ 100 ]	19.7	72.1	30.8	51.1	34.3	47.7	12.7	89.4	11.7	96.9	12.3	91.7	30.1	62.3	29.5	55.4	7.0
MoGe2 [ 100 ] †	7.1	94.6	25.8	60.2	33.1	52.1	7.8	97.5	10.5	98.4	6.5	97.2	8.9	92.1	15.8	84.1	3.9
CUT3R [ 98 ]	8.2	93.6	34.9	45.9	42.9	35.8	6.5	98.0	16.0	88.1	57.9	14.0	17.5	78.3	17.2	81.6	6.9

Table 9: Scale-invariant video pointmap evaluation . Results are aligned with the ground truth by optimizing a shared scale factor across the entire video. We mark best and second-best .

Method	GMU [ 34 ]		Monkaa [ 64 ]		Sintel [ 9 ]		ScanNet [ 20 ]		KITTI [ 33 ]		UrbanSyn [ 37 ]		Unreal4K [ 92 ]		Diode [ 93 ]		Rank $\downarrow$
Method	$\mathrm{Rel}^{d}\!\downarrow$	$\delta^{d}\!\uparrow$	$\mathrm{Rel}^{d}\!\downarrow$	$\delta^{d}\!\uparrow$	$\mathrm{Rel}^{d}\!\downarrow$	$\delta^{d}\!\uparrow$	$\mathrm{Rel}^{d}\!\downarrow$	$\delta^{d}\!\uparrow$	$\mathrm{Rel}^{d}\!\downarrow$	$\delta^{d}\!\uparrow$	$\mathrm{Rel}^{d}\!\downarrow$	$\delta^{d}\!\uparrow$	$\mathrm{Rel}^{d}\!\downarrow$	$\delta^{d}\!\uparrow$	$\mathrm{Rel}^{d}\!\downarrow$	$\delta^{d}\!\uparrow$	Rank $\downarrow$
DepthPro [ 8 ]	8.8	93.0	23.3	55.8	36.1	49.5	8.1	94.6	11.7	93.6	51.3	38.3	105.0	20.0	31.0	58.8	8.0
MoGe [ 99 ]	19.9	66.5	19.9	63.6	26.5	60.0	12.5	85.9	7.6	94.2	15.2	82.1	38.7	46.2	31.3	48.2	7.6
MoGe2 [ 100 ]	19.0	68.9	20.8	60.8	26.4	59.9	12.1	86.9	6.7	96.7	13.8	85.4	33.6	52.8	29.5	50.4	6.8
MoGe2 [ 100 ] †	6.6	93.8	18.0	68.1	25.0	63.8	6.4	96.8	5.2	98.3	7.7	96.4	12.0	88.3	14.7	80.7	3.8
CUT3R [ 98 ]	7.3	92.9	28.0	49.9	31.9	50.5	5.4	97.4	10.2	89.1	48.6	36.7	15.4	79.7	16.0	78.4	6.8

Table 10: Affine-invariant video depthmap evaluation . Results are aligned with the ground truth by optimizing a shared scale and shift factor across the entire video. We mark best and second-best .

Method	GMU [ 34 ]		Monkaa [ 64 ]		Sintel [ 9 ]		ScanNet [ 20 ]		KITTI [ 33 ]		UrbanSyn [ 37 ]		Unreal4K [ 92 ]		Diode [ 93 ]		Rank $\downarrow$
Method	$\mathrm{Rel}^{d}\!\downarrow$	$\delta^{d}\!\uparrow$	$\mathrm{Rel}^{d}\!\downarrow$	$\delta^{d}\!\uparrow$	$\mathrm{Rel}^{d}\!\downarrow$	$\delta^{d}\!\uparrow$	$\mathrm{Rel}^{d}\!\downarrow$	$\delta^{d}\!\uparrow$	$\mathrm{Rel}^{d}\!\downarrow$	$\delta^{d}\!\uparrow$	$\mathrm{Rel}^{d}\!\downarrow$	$\delta^{d}\!\uparrow$	$\mathrm{Rel}^{d}\!\downarrow$	$\delta^{d}\!\uparrow$	$\mathrm{Rel}^{d}\!\downarrow$	$\delta^{d}\!\uparrow$	Rank $\downarrow$
DepthPro [ 8 ]	9.4	92.1	26.7	45.9	53.6	35.3	8.8	92.9	8.2	92.5	22.2	40.2	96.0	1.1	29.3	56.4	7.9
MoGe [ 99 ]	20.7	64.7	25.5	54.8	31.4	48.9	13.3	85.0	7.7	94.1	13.1	86.3	34.7	49.8	29.8	48.2	7.6
MoGe2 [ 100 ]	19.5	67.1	27.1	51.6	31.2	47.7	12.0	86.5	7.2	96.7	12.0	88.8	30.7	56.6	28.5	50.7	6.9
MoGe2 [ 100 ] †	6.7	93.8	21.9	60.4	30.1	51.9	7.1	95.9	5.6	98.3	6.0	96.8	8.7	91.0	14.7	80.7	3.7
CUT3R [ 98 ]	7.9	92.6	33.0	38.3	37.3	42.4	5.8	97.0	11.3	86.8	22.2	63.1	16.8	79.6	15.6	80.0	6.7

Table 11: Scale-invariant video depthmap evaluation . Results are aligned with the ground truth by optimizing a shared scale factor across the entire video. We mark best and second-best .

Method	GMU [ 34 ]		ScanNet [ 20 ]		KITTI [ 33 ]		UrbanSyn [ 37 ]		Diode [ 93 ]
Method	$\mathrm{Rel}^{p}\!\downarrow$	$\delta^{p}\!\uparrow$	$\mathrm{Rel}^{p}\!\downarrow$	$\delta^{p}\!\uparrow$	$\mathrm{Rel}^{p}\!\downarrow$	$\delta^{p}\!\uparrow$	$\mathrm{Rel}^{p}\!\downarrow$	$\delta^{p}\!\uparrow$	$\mathrm{Rel}^{p}\!\downarrow$	$\delta^{p}\!\uparrow$
CUT3R [ 98 ]	13.5	90.7	9.1	95.2	34.2	14.6	15.4	84.4	31.6	47.2
MapAny [ 49 ]	22.6	63.4	35.2	28.5	29.1	26.3	28.5	37.3	33.8	31.8
DAGE (ours)	7.5	95.3	2.5	99.5	12.0	98.3	8.3	96.5	12.9	87.5

Table 12: Metric video pointmap evaluation . Predicted pointmaps are directly compared with ground truth.

Method	GMU [ 34 ]		Monkaa [ 64 ]		Sintel [ 9 ]		ScanNet [ 20 ]		KITTI [ 33 ]		UrbanSyn [ 37 ]		Unreal4K [ 92 ]		Diode [ 93 ]
	( $960\times 512$ )		( $960\times 512$ )		( $896\times 448$ )		( $640\times 512$ )		( $768\times 384$ )		( $2048\times 1024$ )		( $1920\times 1080$ )		( $1024\times 768$ )
	$\mathrm{Rel}^{d}\!\downarrow$	$\delta^{d}\!\uparrow$	$\mathrm{Rel}^{d}\!\downarrow$	$\delta^{d}\!\uparrow$	$\mathrm{Rel}^{d}\!\downarrow$	$\delta^{d}\!\uparrow$	$\mathrm{Rel}^{d}\!\downarrow$	$\delta^{d}\!\uparrow$	$\mathrm{Rel}^{d}\!\downarrow$	$\delta^{d}\!\uparrow$	$\mathrm{Rel}^{d}\!\downarrow$	$\delta^{d}\!\uparrow$	$\mathrm{Rel}^{d}\!\downarrow$	$\delta^{d}\!\uparrow$	$\mathrm{Rel}^{d}\!\downarrow$	$\delta^{d}\!\uparrow$
CUT3R [ 98 ]	22.0	67.9	30.9	51.0	38.9	40.9	7.0	97.9	13.2	88.7	56.7	13.9	71.6	5.6	31.1	52.6
VGGT [ 96 ]	15.9	91.4	17.7	81.6	28.7	63.8	4.5	99.1	7.8	97.5	OOM	OOM	OOM	OOM	20.5	76.3
Pi3 [ 103 ]	6.2	92.2	12.6	88.9	21.7	72.9	2.2	99.5	5.9	97.5	55.9	14.7	54.2	17.1	13.9	87.2
DAGE (ours)	4.9	94.2	10.1	91.0	21.5	75.6	2.1	99.5	5.9	99.0	8.8	96.0	11.9	89.1	9.7	94.4

Table 13: Affine-invariant video pointmap evaluation at native resolution. Predictions are aligned to ground truth by optimizing a single scale and shift across the entire video.

Method	NYUv2 [ 83 ]		KITTI [ 33 ]		ETH3D [ 80 ]		iBims-1 [ 51 ]		GSO [ 25 ]		Sintel [ 9 ]		DDAD [ 35 ]		DIODE [ 93 ]		HAMMER [ 44 ]		Rank $\downarrow$
Method	$\mathrm{Rel}^{d}\!\downarrow$	$\delta^{d}\!\uparrow$	$\mathrm{Rel}^{d}\!\downarrow$	$\delta^{d}\!\uparrow$	$\mathrm{Rel}^{d}\!\downarrow$	$\delta^{d}\!\uparrow$	$\mathrm{Rel}^{d}\!\downarrow$	$\delta^{d}\!\uparrow$	$\mathrm{Rel}^{d}\!\downarrow$	$\delta^{d}\!\uparrow$	$\mathrm{Rel}^{d}\!\downarrow$	$\delta^{d}\!\uparrow$	$\mathrm{Rel}^{d}\!\downarrow$	$\delta^{d}\!\uparrow$	$\mathrm{Rel}^{d}\!\downarrow$	$\delta^{d}\!\uparrow$	$\mathrm{Rel}^{d}\!\downarrow$	$\delta^{d}\!\uparrow$	Rank $\downarrow$
DepthPro [ 8 ]	4.36	97.9	9.15	90.7	7.73	94.0	4.34	97.4	3.16	99.7	19.6	74.5	14.4	81.2	6.28	93.7	5.31	98.8	3.8
MoGe [ 99 ]	3.68	98.3	4.86	97.2	3.57	99.0	3.61	97.3	1.14	100	16.8	77.8	10.5	91.4	4.37	96.4	3.88	98.1	1.8
MoGe2 [ 100 ]	3.33	98.4	6.47	96.4	3.89	98.7	3.65	98.5	1.16	100	17.4	77.0	10.1	90.3	5.13	94.9	4.19	99.1	1.9
DAGE (ours)	3.34	98.4	7.52	94.7	3.49	98.0	3.70	97.8	1.26	99.9	18.9	74.8	10.7	89.2	4.97	94.6	3.43	98.6	2.5

Table 14: Single-image geometry evaluation . Results are aligned with the ground truth by optimizing a scale and shift factor for each image. We mark best and second-best .

Method	RealEstate10K			Co3Dv2
Method	RRA@30 $\uparrow$	RTA@30 $\uparrow$	AUC@30 $\uparrow$	RRA@30 $\uparrow$	RTA@30 $\uparrow$	AUC@30 $\uparrow$
VGGT (518px)	99.97	93.13	77.62	98.64	97.62	91.28
Pi3 (518px)	99.99	95.62	85.90	98.49	97.53	91.39
DAGE (252px)	99.98	95.22	83.12	98.74	97.71	90.71

Table 15: Pose Estimation on RealEstate10K and Co3Dv2

Table 16: Ablations. (a) LR-stream architectures. (b,c) Positional encodings.

Figure 7: Visualization of 3D pointmap reconstruction on in-the-wild scenarios.

Figure 8: Visualization of video depth estimation . We compare our video depth prediction with VGGT [ 96 ] , Pi3 [ 103 ] , and GeoemtryCrafter [ 109 ] . DAGE demonstrates more sharp and fine-grained predictions.

Figure 9: Visualization of video depth estimation . We compare our video depth prediction with VGGT [ 96 ] , Pi3 [ 103 ] , and GeoemtryCrafter [ 109 ] . DAGE demonstrates more sharp and fine-grained predictions.

Figure 10: Visualization of depth estimation on static scenes.

Figure 11: Visualization of predicted depth edge maps , which are defined by a depth ratio between neighboring pixels above a threshold. We zoom-in the edge map details in the red bounding boxes .

Figure 12: Predicted 3D pointmaps of the aligned MoGe2 baseline and our method. The aligned MoGe2 baseline exhibits layering artifacts ( green boxes ) due to the lack of strong multi-view binding.

Figure 13: Visualization of 3D reconstruction with high-resolution inputs.

Panel of Figure 14 — Figure 14: Qualitative results for high-resolution single-image inference: (a) VGGT [ 96 ] and (b) Pi3 [ 103 ] .

Figure 15: Attention map of the 15th global-attention layer of VGGT [ 96 ] at different input resolutions. The query token in the first image is marked with a blue star .

Figure 16: Attention map of the 15th global-attention layer of VGGT [ 96 ] at different input resolutions. The query token in the first image is marked with a blue star .

Figure 17: Attention map of the 15th global-attention layer of VGGT [ 96 ] at different input resolutions. The query token in the first image is marked with a blue star .

Multi-View 3D Reconstruction

Multi-View 3D Reconstruction recovers scene geometry and appearance from images captured at multiple viewpoints.

See all 265 papers

May 28, 2026

Towards Consistent Video Geometry Estimation

This work presents ViGeo, a feed-forward foundation model for recovering spatially dense and temporally consistent geometry from video sequences. Built upon a plain transformer architecture without task-specific architectural modifications, ViGeo supports streaming, full-sequence, and long-video inference within a unified model. The key design is dynamic chunking attention, which exposes the model to both bidirectional and causal temporal contexts during training and allows it to adapt its attention pattern at test time without retraining. To improve supervision quality, we further introduce a completion-based data refinement framework. This framework trains a video depth completion teacher that conditions on sparse and noisy annotations and exploits video/multi-view context to produce dense, temporally coherent, and geometrically reliable training targets. Beyond depth and point maps, ViGeo also predicts surface normals within the same framework. Trained solely on public datasets, ViGeo achieves state-of-the-art performance across online, offline, and long-video depth estimation, surface normal estimation, and video point map estimation.

Zhu Yu, Jingnan Gao, Runmin Zhang +9

Zhejiang University · Shanghai Jiao Tong University · Tongyi Lab, Alibaba Group +1

Vision Foundation Models Depth Estimation Surface Normal Estimation Video Depth Estimation

Ask PDF

Jul 2, 2026

Diversity-aware View Partitioning for Scalable VGGT

Geometry transformers such as VGGT achieve strong performance by jointly reasoning over multiple views with global attention. However, scaling them to large view collections remains challenging due to the quadratic cost of attention. Moreover, our empirical analysis reveals that the reconstruction quality in VGGT is sensitive to the distribution of viewpoints. Simply increasing the number of views without sufficient viewpoint diversity can even degrade performance, as redundant views introduce highly similar tokens that dilute informative geometric signals in the attention mechanism. Motivated by this observation, we propose a training-free and plug-and-play VGGT inference framework that organizes views into diversity-aware balanced chunks. The chunks are constructed through combinatorial graph partitioning over visual dissimilarity and spatial dispersion. This view organization allows the transformer to focus attention on geometrically informative views while reducing redundant attention interactions. To estimate spatial dispersion without full pose estimation, we approximate spatial relationships via a soft pose propagation strategy based on visual similarity from a small set of seed frames. Extensive experiments demonstrate improved performance in camera pose estimation, multi-view depth prediction, and 3D reconstruction while reducing memory usage and inference latency. Our framework also complements existing VGGT variants, enabling scalable multi-view reconstruction without sacrificing geometric fidelity.

Jinsoo Park, Donggyu Choi, Ahyun Seo +2

POSTECH · GIST · KAIST +1

3D ViTs Multi-View 3D Reconstruction Camera Pose Estimation Efficient ViTs

Ask PDF

Jun 16, 2026

RegimeVGGT: Layer-Wise Spatially Preserving Redundancy Removal for Visual Geometry Grounded Transformer

Visual Geometry Grounded Transformer (VGGT) recovers dense 3D scene structure from multi-view images in one forward pass, but quadratic cross-frame attention limits its scalability. Existing training-free accelerators reduce computation uniformly along one axis, missing layer heterogeneity. Our spectral, probing, and causal analyses reveal three regimes: shallow layers lack cross-view structure, middle layers drive cross-view alignment, and deep layers are redundant for dense geometry yet their cross-frame attention remains essential for pose. RegimeVGGT applies layer-wise U-shaped compression along two axes: Saliency-Guided Banded Merging protects geometry- and edge-salient tokens, while Selectively Protected K/V Downsampling preserves cross-frame spatial coverage and the pose-critical path through a phase-shifted spatial grid, a reference-frame anchor, and uncompressed camera/register tokens. Training-free, RegimeVGGT achieves a 6.7x speedup over VGGT* at matched reconstruction quality.

Jinhao You, Shuo Lyu, Zhuohang Lyu +5

University of Pennsylvania · University of California, Irvine · Nanyang Technological University +1

3D ViTs Multi-View 3D Reconstruction Efficient ViTs Visual Token Compression

Ask PDF

Abstract

Multi-View 3D Reconstruction

Explore similar work

Towards Consistent Video Geometry Estimation

Diversity-aware View Partitioning for Scalable VGGT

RegimeVGGT: Layer-Wise Spatially Preserving Redundancy Removal for Visual Geometry Grounded Transformer