Trainspotting: Learning Object-Centric 3D Reconstruction from Found Photos

Hanyu Chen; Noah Snavely

cs.CVNov 4, 2026

Trainspotting: Learning Object-Centric 3D Reconstruction from Found Photos

Authors: Hanyu Chen, Noah Snavely

Organizations: Cornell University

Abstract

State-of-the-art feedforward reconstruction models perform poorly on in-the-wild object-centric photo collections. They are trained largely on single-session object captures whose images vary in viewpoint while the object and its environment remain nearly fixed. Public photo archives contain a different form of multi-view data. They hold found photos of the same object or design: photos taken independently by many people across years, environments, cameras, and sometimes physical units. In this work, we introduce Trainspotting, a dataset of 12,068 objects and 311,886 photos from Wikimedia Commons. We reconstruct each object to recover camera poses, multi-view stereo depth, and object masks. We categorize the objects into three regimes: objects fixed at one site, objects that move between sites, and designs shared by many physical units. Evaluating recent feedforward reconstruction models on Trainspotting shows that they fail most severely in the last two regimes. Fine-tuning VGGT- $Ω$ on Trainspotting significantly improves camera pose, depth, and 3D geometry on held-out objects. We observe similar improvements on the independent NAVI in-the-wild benchmark, which contains photos of each object in different environments. Performance on conventional benchmarks remains largely unchanged.

Figures & tables

Figure 2: VGGT- $\Omega$ fails on the Buick Riviera . We visualize the camera poses and point map predicted by VGGT- $\Omega$ from sixteen photos of the Buick Riviera (multiple physical instances of the same car design). The views are misregistered and the car is reconstructed as scattered fragments. Figure 6 shows the reconstruction of our fine-tuned model from the same photos.

Figure 3: Examples of capture variation. Each column displays photos of example objects, grouped into the three regimes (fixed, roaming, and instance-pooled) and labeled with their year and camera. Fixed objects stay at one site, roaming objects are single instances that move between sites, and instance-pooled objects are different physical instances of one shared design.

Dataset	Type	Count	Images	Span	Depth
OmniObject3D [ 53 ]	objects	6k	1.2M	minutes	–
CO3D [ 31 ]	objects	19k	1.5M	minutes	MVS
MVImgNet2.0 [ 54 ]	objects	520k	15.5M	minutes	–
uCO3D [ 28 ]	objects	170k	34M	minutes	mono.
NAVI [ 19 ]	objects	36	10.5k	days	scan
Pix3D [ 40 ]	objects	395	10k	days/years	scan/CAD

Table 1: Real-image datasets for 3D reconstruction tasks. Datasets above the line record each object or scene in a single capture within minutes, while datasets below contain photos of each object or scene taken by many people spanning days to years. NAVI and Pix3D align images to 3D models by hand; all other datasets recover poses with structure from motion.

Figure 4: Reconstruction pipeline for the Flying Scotsman. (a) Object gating retains 570 of 727 photos, with red borders marking excluded photos. (b) A retrieved pair is shown with candidate correspondences. Green correspondences pass the confidence, cycle, and epipolar checks, while dashed red correspondences fail. (c) The structure-from-motion reconstruction registers 496 images, merged from 14 partial reconstructions. View selection keeps 481 of them for multi-view stereo. (d) The same pair shows MVS depth before and after MoGe-2 filtering. (e) Four of the 370 photos in the dataset show MVS depth in color and the selected mask in teal over the image.

Supercategory	Objects	Supercategory	Objects
Road vehicles	3,341	Household & personal objects	353
Statues & sculpture	2,160	Weapons & armour	342
Monuments & landmarks	1,773	Religious objects	140
Aircraft & spacecraft	1,461	Street furniture	130
Rail vehicles	1,108	Machinery & equipment	121
Watercraft	701	Other	10

Table 2: Dataset composition. Number of objects in each supercategory. An LLM groups the object labels assigned during curation into 53 categories and 13 supercategories.

Figure 5: Object statistics by regime. Distributions over Trainspotting of the number of photos, the span of capture dates, the number of photographers, and the number of camera models.

Figure 6: Qualitative comparison. We visualize ground truth and predicted camera poses and point maps from $\pi^{3}$ [ 52 ] , Depth Anything 3 [ 25 ] , VGGT- $\Omega$ [ 49 ] , and Spotter on held out objects from Trainspotting and NAVI in-the-wild [ 19 ] . We use 4 or 16 input views for Trainspotting and 8 for NAVI in-the-wild. Only the first 4 input views are shown. Triangular markers indicate the directions of the cameras outside of the frame. The baselines misregister the input views and produce duplicated or fragmented geometry, whereas Spotter recovers a single consistent object and accurate camera poses.

Method	Point map		Depth		Camera
	Acc $\downarrow$	Comp $\downarrow$	AbsRel $\downarrow$	$\delta_{1.03}\uparrow$	AUC ${}_{5}\uparrow$	AUC ${}_{30}\uparrow$
VGGT [ 48 ]	2.38	8.08	0.160	32.4	11.5	44.1
$\pi^{3}$ [ 52 ]	2.33	6.22	0.162	34.7	12.5	46.4
DA3 [ 25 ]	2.22	12.17	0.189	32.1	10.1	29.4
MapAny. [ 20 ]	2.20	7.75	0.206	26.6	5.3	33.1
VGGT- $\Omega$ [ 49 ]	1.84	12.68	0.201	28.1	7.2	25.0

Table 3: Results on Trainspotting . Point-map accuracy and completeness are given as a percentage of the object extent after a similarity alignment refined by ICP. Depth AbsRel and inlier ratio

\delta_{1.03}

are measured after a single median scale per sample. Camera-pose AUC at

5^{\circ}

and

30^{\circ}

is computed over all pairs of views (Section 5.2 ).

Method	NAVI (wild)			NAVI			7-Scenes			NRGBD			DTU			ETH3D
	Acc $\downarrow$	Comp $\downarrow$	$\delta_{1.03}\uparrow$	Acc $\downarrow$	Comp $\downarrow$	$\delta_{1.03}\uparrow$	Acc $\downarrow$	Comp $\downarrow$	$\delta_{1.03}\uparrow$	Acc $\downarrow$	Comp $\downarrow$	$\delta_{1.03}\uparrow$	Acc $\downarrow$	Comp $\downarrow$	$\delta_{1.03}\uparrow$	Acc $\downarrow$	Comp $\downarrow$	$\delta_{1.03}\uparrow$
VGGT [ 48 ]	3.2	11.5	35.0	2.5	3.4	71.5	25.3	32.6	67.5	17.4	16.9	90.9	1.2	1.8	97.4	291.4	317.5	52.6
$\pi^{3}$ [ 52 ]	3.7	9.6	41.2	2.2	3.0	76.5	20.0	26.7	70.5	17.9	17.2	88.6	1.2	1.5	91.2	154.7	129.7	62.4
DA3 [ 25 ]	3.3	16.0	41.1	3.3	5.7	80.0	25.4	35.7	74.1	16.4	15.3	96.7	1.7	1.6	96.5	274.3	270.3	73.2
MapAny. [ 20 ]	4.1	11.0	29.6	3.8	5.8	56.4	28.2	36.8	58.3	42.8	30.2	69.2	2.4	1.9	37.8	459.7	542.0	51.9
VGGT- $\Omega$ [ 49 ]	3.1	20.4	29.7	2.9	5.3	76.1	21.8	27.4	71.9	13.3	9.0	92.4	1.2	1.7	97.1	73.6	61.1	83.5

Table 5: Geometry on external benchmarks. Point-map accuracy and completeness are given in millimeters after a similarity alignment refined by ICP. Depth inlier ratio

\delta_{1.03}

is measured after a single median scale per sample. Every model is given the same 8 views of each scene. NAVI photographs each object in one session ( NAVI ) and across different environments ( NAVI (wild) ). The other datasets record each scene in a single capture.

Training data	All	Fixed	Roaming	Pooled	NAVI (wild)
	Trainspotting held-out
All (11,068)	80.3	78.2	79.5	82.1	74.6
Random 20% (2,193)	79.7	77.5	79.2	81.3	72.3
Fixed (4,038)	68.5	78.4	58.0	68.9	63.4
Roaming (2,603)	79.0	76.5	80.3	79.7	65.7
Pooled (4,427)	78.8	75.6	77.0	81.7	73.9

Table 7: Training-data ablations. AUC@30 of models trained for 10 epochs on subsets of the Trainspotting training objects. The number of training objects is given in parentheses.

Appendix figures & tables11 assets

Supplementary material from the paper’s appendix.

Appendix

Category	Objects	Photos	Category	Objects	Photos
Road vehicles			Musical instruments
car	1,967	87,193	pipe organ	390	5,733
bus	476	28,932	musical instrument	38	982
military vehicle	341	7,522	Household & personal objects
truck & van	269	8,553	camera & lens	142	2,946
motorcycle	187	2,776	computer & office machine	90	1,364

Appendix

Table S1 : Object categories. The 53 categories, grouped by supercategory, with the number of objects and photos of each (12,068 objects and 311,886 photos in total).

Reason	Categories
Flat object, such as a painting or a sign	1,802
Media or two-dimensional depiction of an object	1,606
Place or institution	1,489
Abstract or maintenance category	1,066
Event or activity	811
Deformable object	767

Appendix

Table S2 : Categories rejected by the classifier. Reasons for the 8,956 categories with at least 25 photos that the LLM does not keep.

Regime	Objects	MegaScenes	MegaScenes	MegaDepth-X
		category	reconstruction
Fixed	4,338	1,798	1,128	56
Roaming	2,869	41	34	0
Instance-pooled	4,861	2	1	0
Total	12,068	1,841	1,163	56

Appendix

Table S3 : Overlap with internet-photo scene datasets by regime. Number of Trainspotting objects whose Wikimedia Commons category is a MegaScenes scene, has a released MegaScenes reconstruction, or is a MegaDepth-X scene.

Stage	Objects	Photos
Input	17,470	1,251,986
Object gating	16,141	775,609
Structure from motion	14,818	558,281
MVS depth	14,032	486,275
Mask and depth checks	13,236	367,245
Final dataset	12,068	311,886

Appendix

Table S4 : Reconstruction yield. Number of objects and photos that remain after each stage of the pipeline in Section 4 of the main paper.

Figure S1 : Structure from motion with other matchers. Each row shows one Trainspotting object: six of the input photos, then the largest model of COLMAP with SIFT, SuperPoint with LightGlue, MASt3R matches with the COLMAP mapper, MASt3R-SfM, and our pipeline, all run on the same images. The number of registered images out of the input images is given under each panel. The other methods fail on instance-pooled and roaming objects, and on fixed objects our pipeline produces a more complete and cleaner reconstruction. In the bottom three rows, we frame the object to show its details, so not all cameras are visible.

Stage	Threshold	Value	Definition
Object gating	minimum box size	10% of the image area	Largest SAM3 detection, with detection score above 0.5.
Retrieval	neighbours per image	20
Matching	correspondences per pair	8,000	Sampled from the dense warp.
	$\tau_{\sigma}$	0.20	Certainty threshold.
	$\tau_{\mathrm{cyc}}$	0.4% of the image side	Cycle-consistency distance.
	$\tau_{\mathrm{epi}}$	0.1% of the image side	Epipolar distance.

Appendix

Table S5 : Pipeline thresholds. The fixed thresholds of the reconstruction pipeline of Section 4 of the main paper, in the order of its stages. Image-relative values are fractions of the image side.

Method	Median	10th pct.	90th pct.	Objects below 0.7
VGGT- $\Omega$	0.43	0.15	0.87	28
Spotter	0.92	0.81	0.96	0

Appendix

Table S6 : Extent of the aligned prediction on NAVI in-the-wild. Ratio of the extent of the aligned prediction to the extent of the ground truth over the 35 objects, and the number of objects with a ratio below 0.7.

Figure S2 : Failure cases. Four held-out samples on which Spotter fails: airliner (noisy depth at long focal lengths), car (misregistered close-up of a wheel), grave monument (doppelganger ambiguity), and toy shark from NAVI in-the-wild (out of distribution).

Figure S3 : Additional qualitative comparisons. Ten held-out samples drawn at random: three Trainspotting objects with 16 input views, five with 4 input views, and two NAVI in-the-wild objects with 8 input views.

Method	Point map				Depth				Camera
	Acc mean	Acc med.	Comp mean	Comp med.	AbsRel	$\delta_{1.03}$	$\delta_{1.05}$	$\delta_{1.10}$	AUC 5	AUC 15	AUC 30
VGGT [ 48 ]	2.38	1.62	8.08	6.48	0.160	32.4	41.8	56.2	11.5	29.9	44.1
$\pi^{3}$ [ 52 ]	2.33	1.61	6.22	4.76	0.162	34.7	43.8	56.8	12.5	31.9	46.4
DA3 [ 25 ]	2.22	1.60	12.17	10.11	0.189	32.1	41.5	54.6	10.1	21.0	29.4
MapAny. [ 20 ]	2.20	1.57	7.75	6.18	0.206	26.6	35.3	49.1	5.3	19.3	33.1
VGGT- $\Omega$ [ 49 ]	1.84	1.36	12.68	10.77	0.201	28.1	37.1	50.3	7.2	16.9	25.0

Appendix

Table S7 : All metrics on Trainspotting . Extends Table 3 of the main paper with the median of accuracy and completeness, the inlier ratios at 1.05 and 1.10, and AUC at

15^{\circ}

. Accuracy and completeness are in percent of the object extent.

7-Scenes (18 samples)
Method	Point map					Depth				Camera
	Acc mean	Acc med.	Comp mean	Comp med.	NC	AbsRel	$\delta_{1.03}$	$\delta_{1.05}$	$\delta_{1.10}$	AUC 5	AUC 15	AUC 30
VGGT [ 48 ]	25.3	11.6	32.6	15.9	0.735	0.277	67.5	83.4	92.3	26.6	65.2	81.0
$\pi^{3}$ [ 52 ]	20.0	10.5	26.7	14.3	0.765	0.342	70.5	87.1	96.4	28.8	68.5	83.3
DA3 [ 25 ]	25.4	13.1	35.7	15.9	0.755	0.324	74.1	88.1	94.6	28.5	66.3	81.5
MapAny. [ 20 ]	28.2	14.7	36.8	16.0	0.739	0.332	58.3	77.3	93.3	18.2	54.7	72.4

Appendix

Table S8 : All metrics on the external benchmarks. The metrics of Table S7 and normal consistency (NC) on the benchmarks of Tables 5 and 6 of the main paper. Accuracy and completeness are in millimeters.

Object-Centric 3D Reconstruction

See all 19 papers

VGG-T $^3$ : Offline Feed-Forward 3D Reconstruction at Scale
Feb 26, 2026Sven Elflein, Ruilong Li, Sérgio Agostinho +43D Reconstruction Feed-Forward 3D Reconstruction
NVIDIA · Vector Institute · University of Toronto
OTT3R: Multi-View 3D Reconstruction and Fast Dataset Generation at 1% Compute
Sep 28, 2026Brandon Leblanc, Charalambos PoullisMulti-View 3D Reconstruction Pseudo-Labeling
Immersive and Creative Technologies Lab, Concordia University, Montreal, Canada
Any3R: Feedforward 3D Reconstruction from Any Camera Type
Oct 26, 2026Jingong Chen, Qingwen Zhang, Youming Deng +53D Reconstruction Multi-View 3D Reconstruction
South Dakota State University · Aalto University · KTH Royal Institute of Technology +3

Feb 26, 2026J/K move · Enter open · S save

VGG-T $^3$ : Offline Feed-Forward 3D Reconstruction at Scale

Sven Elflein, Ruilong Li, Sérgio Agostinho, Zan Gojcic+3

NVIDIA · Vector Institute · University of Toronto

3D Reconstruction Feed-Forward 3D Reconstruction Large-Scale 3D Reconstruction 3D Scene Representation

Open paper AskPDF arXiv

We present a scalable 3D reconstruction model that addresses a critical limitation in offline feed-forward methods: their computational and memory requirements grow quadratically w.r.t. the number of input images. Our approach is built on the key insight that this bottleneck stems from the varying-length Key-Value (KV) space representation of scene geometry, which we distill into a fixed-size Multi-Layer Perceptron (MLP) via test-time training. VGG-T

^3

(Visual Geometry Grounded Test Time Training) scales linearly w.r.t. the number of input views, similar to online models, and reconstructs a

1k

image collection in just

54

seconds, achieving a

11.6\times

speed-up over baselines that rely on softmax attention. Since our method retains global scene aggregation capability, our point map reconstruction error outperforming other linear-time methods by large margins. Finally, we demonstrate visual localization capabilities of our model by querying the scene representation with unseen images.

Abstract

Object-Centric 3D Reconstruction

Explore similar work

VGG-T3^33: Offline Feed-Forward 3D Reconstruction at Scale

OTT3R: Multi-View 3D Reconstruction and Fast Dataset Generation at 1% Compute

Any3R: Feedforward 3D Reconstruction from Any Camera Type

VGG-T $^3$ : Offline Feed-Forward 3D Reconstruction at Scale