MegaDepth-X: Long-Tail 3D Photo Reconstruction

• MegaDepth-X is a large-scale dataset for long-tail Internet photo 3D reconstruction, offering clean, dense depth maps and refined multi-view supervision.
• It extends the original MegaDepth by incorporating doppelganger pruning, monocular-depth-guided filtering, and a sparse-subset sampling strategy tailored for weakly connected scenes.
• The framework enables effective fine-tuning of 3D foundation models, yielding improved reconstruction metrics and robustness in sparse, challenging scenarios.

Searching arXiv for MegaDepth-X and related papers to ground the article in current literature. MegaDepth-X (MD-X) is a large-scale dataset of Internet-photo 3D reconstructions with clean, dense depth, together with a strategy for sampling sets of training images that mimic camera distributions in long-tail scenes. It is introduced for long-tail Internet photo reconstruction, where photo collections are sparse, noisy, weakly connected, and often outside the operating regime of both classical SfM/MVS pipelines and recent feed-forward 3D models. In the formulation of its authors, MD-X is not merely a larger successor to MegaDepth; it is a cleaned, depth-refined subset of MegaScenes paired with a long-tail simulation recipe intended for fine-tuning 3D foundation models under extreme sparsity while preserving performance on standard dense-scene benchmarks (Li et al., 24 Apr 2026).

1. Long-tail Internet photo reconstruction as the target regime

The motivating observation behind MegaDepth-X is that Internet photo collections are extremely long-tailed. In MegaScenes, the “head” contains 6,985 scenes with more than 50 registered images, whereas the “tail” contains 418,056 scenes with fewer than 50 registered photos. This establishes that the dominant regime is not the densely photographed landmark collection traditionally favored by reconstruction systems, but the sparse, weakly constrained collection in which overlap is limited, viewpoints are uneven, and nuisance content is common (Li et al., 24 Apr 2026).

The paper characterizes the long tail through failure modes that are already familiar in web-scale 3D: sparse or weak overlap, wide baselines, uneven viewpoint distribution, noisy or irrelevant photos, dynamic objects / crowds, and doppelganger ambiguity. In this setting, classical pipelines such as COLMAP may register few images or none at all, and MVS often lacks sufficient support for reliable dense geometry. Recent learned systems such as DUSt3R, VGGT, and $\pi^3$ are described as having strong priors but being trained mainly on controlled captures with clean, dense, evenly sampled observations, which leaves them exposed when the input collection becomes weakly connected or fragmented (Li et al., 24 Apr 2026).

The paper supports this claim with view-graph statistics. In low-registration-rate scenes, 8\% of cameras have degree $\le 2$, compared with 3\% in well-reconstructed head scenes, and the average number of geometrically verified feature matches per connected image pair is 294.8 in long-tail scenes versus 395.3 in head scenes. This suggests that the long tail is not simply a smaller version of the head; it is a distinct observation regime with different graph structure, weaker local reconstructability, and more severe ambiguity. MegaDepth-X is designed around that distinction.

2. Dataset definition, scale, and relationship to MegaDepth

MegaDepth-X is presented as a “next-generation extension” of MegaDepth. The original MegaDepth introduced Internet-photo-derived supervision for single-view depth by reconstructing landmark collections with SfM and MVS, then cleaning dense depth maps and supplementing them with ordinal supervision where reconstruction failed (Li et al., 2018). MegaDepth-X inherits the premise that Internet photos can serve as scalable geometric supervision, but reorients the data pipeline toward multi-view Internet photo reconstruction and sparse-view 3D model fine-tuning rather than scale-invariant monocular depth learning (Li et al., 24 Apr 2026).

Its construction begins from 2,474 candidate scenes from MegaScenes with more than 100 registered images. After filtering out 609 scenes, the final MD-X dataset contains 1,865 reconstructions and 440k images, with a test set of 127 scenes. The paper does not explicitly state the total number of depth maps, nor does it provide a formal dataset schema listing filenames, serialization formats, or exact metadata fields. What is explicitly stated is that MD-X contains Internet photo collections, reliable sparse reconstructions, dense depth maps derived and refined from MVS, camera poses, and scene-graph or view-graph structure sufficient to support graph-based subset sampling (Li et al., 24 Apr 2026).

A concise comparison with MegaDepth is useful because the difference is methodological as well as quantitative.

Aspect	MegaDepth	MegaDepth-X
Reconstructions	266	1,865
Images	119k	440k
Doppelganger check	no	yes
Dense depth refinement	yes	yes, with monocular-depth-guided filtering

The paper also reports broader camera coverage for MD-X. For positional azimuth coverage, the number of scenes with at least 75\% coverage rises from 15 in MegaDepth to 80 in MD-X, and the number with at least 25\% coverage rises from 74 to 752. For rotational azimuth coverage, the number of scenes with at least 75\% coverage rises from 56 to 490, and with at least 25\% coverage from 230 to 1816. The intended significance is that MD-X is better suited to sampling sparse but still meaningful subsets from within a richly reconstructed scene.

A common misconception is to treat MegaDepth-X as simply “more MegaDepth.” The paper argues for a narrower and more specific interpretation: MegaDepth is described as clean but small, whereas MD-X adds scale, doppelganger-aware cleaning, monocular-depth-guided refinement, and a sparse-subset sampling framework specifically tuned to emulate long-tail Internet photo collections.

3. Reconstruction pipeline and dense-depth refinement

MegaDepth-X is built by combining automated reconstruction, explicit filtering, and manual validation. The process begins with scene selection from MegaScenes, restricted to scenes with more than 100 registered images. This is a deliberate decision: the paper does not attempt to reconstruct truly long-tail scenes directly for supervision, because those are often too unreliable to support dense or even sparse geometric ground truth. Instead, it starts from the head of the distribution and simulates the tail during training (Li et al., 24 Apr 2026).

The cleaning pipeline has three major stages. First, scenes dominated by dynamic content are manually excluded. Second, the sparse reconstruction is recomputed with MASt3R-SfM and a Doppelganger classifier from Doppelgangers++, rather than inherited from default COLMAP output. The MASt3R-SfM stage builds a scene graph using feature matches from MASt3R descriptors, while the doppelganger classifier prunes suspicious edges likely induced by false visual correspondences. Third, reconstructed scenes are manually verified against Google Maps and satellite imagery, and scenes not aligned with those external references are discarded (Li et al., 24 Apr 2026).

Dense depth generation then proceeds through a standard MVS stage followed by refinement inherited from MegaDepth and extended by MD-X. The MegaDepth-style refinement includes a modified MVS procedure that conservatively retains the minimum depth value during propagation, stability filtering to remove flickering pixels, and semantic filtering to exclude transient objects. The original MegaDepth paper already established the importance of aggressive cleaning over naive use of raw MVS outputs, showing large cross-dataset improvements when raw MVS depths were replaced by cleaned supervision (Li et al., 2018).

MegaDepth-X adds a further monocular-depth-guided filtering stage using MoGe2 as an ordinal prior. Geometric depth $D_\text{geom}$ is first scale-aligned to monocular depth $D_\text{mono}$ over valid pixels $P$:

$$D'_\text{geom}(p) = s \cdot D_\text{geom}(p), \quad \text{where } s = \frac{\text{med}\{D_\text{mono}(p)\mid p \in P\}}{\text{med}\{D_\text{geom}(p)\mid p \in P\}}.$$

A normalized discrepancy is then computed:

$$\Delta(p) = \frac{\left|D'_\text{geom}(p) - D_\text{mono}(p)\right|}{D'_\text{geom}(p)}.$$

Pixels whose discrepancy exceeds $\tau_{\text{depth}}$ are discarded. A gradient discrepancy is also used:

$$\Delta(p)_{\text{grad}} = \left| \frac{|\nabla D_\text{mono}|}{D_\text{mono}} - \frac{|\nabla D'_\text{geom}|}{D'_\text{geom}} \right|.$$

Pixels whose gradient discrepancy exceeds $\tau_{\text{grad}}$ are likewise discarded (Li et al., 24 Apr 2026).

The paper is explicit that monocular depth is used only as guidance, not as warped pseudo-depth. It prefers accurate depth maps over complete ones, and does not fill missing regions from the monocular predictor because relative monocular depth uncertainty could create cross-view inconsistency. The exact values of $\le 2$0 and $\le 2$1 are not given in the supplied text, nor are camera parameter storage format, world-coordinate convention, depth units, or confidence-mask file format.

4. Long-tail simulation through graph-based sparse-subset sampling

A central contribution of MegaDepth-X is the claim that reliable supervision for the long tail can be simulated by sampling sparse subsets from well-reconstructed scenes. The paper frames this as a solution to the supervision paradox of long-tail Internet reconstruction: truly sparse scenes are often too poorly reconstructed to supervise learning, but richly reconstructed scenes can be subsampled to mimic the same graph statistics (Li et al., 24 Apr 2026).

The scene is represented as a weighted view graph

$\le 2$2

where nodes are cameras and edge weights are feature-match counts $\le 2$3. Weak edges are pruned with

$\le 2$4

yielding a filtered graph $\le 2$5. Communities are then identified with Louvain, and the graph is partitioned into up to $\le 2$6 connected components using parallel round-robin BFS from random seeds. To ensure that sampled batches remain trainable, the method balances three properties: Viewpoint Diversity, Sparsity, and Local Reconstructability (Li et al., 24 Apr 2026).

For each training batch, one representative view is selected from each community, producing a terminal set

$\le 2$7

An approximate Steiner tree is then computed to connect these terminals:

$\le 2$8

Because the Steiner subgraph may still be too large, final views are chosen greedily. If $\le 2$9 is the already sampled set and $D_\text{geom}$0 the node-to-community map, then for each candidate neighbor $D_\text{geom}$1, the method computes community novelty

$D_\text{geom}$2

where

$D_\text{geom}$3

and spatial distance

$D_\text{geom}$4

Candidates are ranked lexicographically by $D_\text{geom}$5 in descending order (Li et al., 24 Apr 2026).

The paper defines four metrics to characterize the effect of the search depth parameter $D_\text{geom}$6. The $D_\text{geom}$7-hop graph coverage is

$D_\text{geom}$8

the nearest-sample distance is

$D_\text{geom}$9

the graph dispersion is

$D_\text{mono}$0

and the Euclidean dispersion is

$D_\text{mono}$1

The reported trend is that larger $D_\text{mono}$2 increases graph coverage and also increases dispersion, producing broader baselines and sparser subsets (Li et al., 24 Apr 2026).

Offline batch generation precomputes 24-node subsets per scene, and training then subsamples 2 to 24 images from those cached batches. The paper defines four concrete regimes: Dense with $D_\text{mono}$3, $D_\text{mono}$4; Sparse with $D_\text{mono}$5, $D_\text{mono}$6; Mixed with $D_\text{mono}$7, $D_\text{mono}$8; and Random with random view sampling. The default fine-tuned model uses Mixed.

5. Fine-tuning 3D foundation models on MegaDepth-X

MegaDepth-X is used primarily as a fine-tuning resource for two feed-forward 3D models: $D_\text{mono}$9 and VGGT, producing $P$0-FT and VGGT-FT. The paper states that it adopts the original loss functions of those backbones rather than introducing new MD-X-specific losses. The supervision comes from the cleaned sparse reconstructions, camera poses, and refined dense depth or point-map information supplied by MD-X (Li et al., 24 Apr 2026).

The fine-tuning protocol is deliberately conservative. To preserve pretrained geometric fidelity, the authors fine-tune only the Alternating-Attention modules and freeze the point cloud and camera decoders. This suggests that MD-X is intended less as a full retraining corpus than as a distributional adaptation mechanism that changes multi-view reasoning under sparsity without discarding existing decoding priors (Li et al., 24 Apr 2026).

The reported training setup is specific. Images are padded to $P$1. Augmentations include random crops with aspect ratio sampled uniformly from $P$2, random color jitter, and random image rotations by $P$3 clockwise or counterclockwise with probability 0.2. Each minibatch contains up to 24 images, the number of views per batch is sampled from [2, 24], and there are at most 96 images per GPU. Training runs for 100 epochs with AdamW, a peak learning rate of $P$4, and a linear warmup + cosine annealing schedule on 4 NVIDIA A6000 GPUs. The fine-tuning mixture also includes BlendedMVS and TartanAir (Li et al., 24 Apr 2026).

The key distinction from standard dense-scene training is the sampling distribution rather than the backbone architecture. Standard training typically uses highly overlapping image sets with strong covisibility; MD-X deliberately feeds sparse subsets, wide baselines, multiple weakly connected components, and mixtures of dense and sparse batches. In that sense, the paper positions MD-X as a data engine for adapting general 3D models to Internet-photo sparsity.

6. Empirical behavior, benchmark role, and limitations

On the MD-X test set, evaluation uses 24 images per scene under an easy regime with $P$5, $P$6, and a hard regime with $P$7, $P$8. Metrics include RRA@5, RTA@5, AUC@5, MRE, and MTE for camera pose, plus Accuracy (Acc), Completeness (Comp), and Normal Consistency (NC) for point maps (Li et al., 24 Apr 2026).

For $P$9 on the easy split, the pretrained model achieves RRA@5 = 88.97, RTA@5 = 68.79, AUC@5 = 45.84, MRE = 4.12, MTE = 7.82, Acc mean = 0.055, Comp mean = 0.039, and NC mean = 0.712. After fine-tuning, $$D'_\text{geom}(p) = s \cdot D_\text{geom}(p), \quad \text{where } s = \frac{\text{med}\{D_\text{mono}(p)\mid p \in P\}}{\text{med}\{D_\text{geom}(p)\mid p \in P\}}.$$0-FT reaches RRA@5 = 95.64, RTA@5 = 76.85, AUC@5 = 55.58, MRE = 1.64, MTE = 5.50, Acc mean = 0.035, Comp mean = 0.024, and NC mean = 0.724. On the hard split, the corresponding AUC@5 improves from 36.93 to 47.93, and Acc mean improves from 0.101 to 0.068 (Li et al., 24 Apr 2026).

For VGGT, the paper reports analogous gains. On easy scenes, VGGT improves from AUC@5 = 35.32 to 48.78 after fine-tuning, with Acc mean dropping from 0.093 to 0.050 and NC mean rising from 0.695 to 0.719. On hard scenes, AUC@5 improves from 29.10 to 41.49, Acc mean from 0.149 to 0.089, and NC mean from 0.675 to 0.709. The largest gains occur under sparse sampling, which is consistent with the paper’s central thesis that long-tail performance depends on training under long-tail-like graph structure (Li et al., 24 Apr 2026).

The paper also presents qualitative evidence on genuinely difficult Internet scenes where COLMAP registers few or no images, including examples with 66 images / 13 registered, 95 images / 11 registered, 69 images / 11 registered, 44 images / 15 registered, and 94 images / 0 registered. The fine-tuned model is described as recovering denser and more coherent geometry, handling doppelganger ambiguity more reliably, and maintaining confidence under severe sparsity (Li et al., 24 Apr 2026).

Ablations reinforce two points. First, clean supervision matters: on hard scenes, pretrained $$D'_\text{geom}(p) = s \cdot D_\text{geom}(p), \quad \text{where } s = \frac{\text{med}\{D_\text{mono}(p)\mid p \in P\}}{\text{med}\{D_\text{geom}(p)\mid p \in P\}}.$$1 has AUC@5 = 36.93, $$D'_\text{geom}(p) = s \cdot D_\text{geom}(p), \quad \text{where } s = \frac{\text{med}\{D_\text{mono}(p)\mid p \in P\}}{\text{med}\{D_\text{geom}(p)\mid p \in P\}}.$$2-Dirty reaches 43.74, and $$D'_\text{geom}(p) = s \cdot D_\text{geom}(p), \quad \text{where } s = \frac{\text{med}\{D_\text{mono}(p)\mid p \in P\}}{\text{med}\{D_\text{geom}(p)\mid p \in P\}}.$$3-FT reaches 47.93, but the dirty variant harms point-map quality, with Acc mean = 0.130 versus 0.101 for pretrained and 0.068 for FT. Second, the sampling strategy matters: Random gives reasonable pose accuracy but weaker point-map gains, Dense favors easier scenes, Sparse exposes difficult cases but is not the best overall trade-off, and Mixed gives the best overall balance (Li et al., 24 Apr 2026).

The paper also argues that MD-X fine-tuning does not severely damage generalization. For $$D'_\text{geom}(p) = s \cdot D_\text{geom}(p), \quad \text{where } s = \frac{\text{med}\{D_\text{mono}(p)\mid p \in P\}}{\text{med}\{D_\text{geom}(p)\mid p \in P\}}.$$4, AUC@5 on RealEstate10K shifts from 62.82 to 60.01, while on CO3Dv2 it shifts from 57.12 to 57.61. For VGGT, RealEstate10K improves from 38.09 to 48.23, and CO3Dv2 remains essentially unchanged at 67.84 versus 67.81. On clean benchmarks such as ETH3D, however, the paper reports some degradation and attributes it to domain mismatch with controlled imagery (Li et al., 24 Apr 2026).

Several limitations are explicit. MegaDepth-X still focuses on landmark-scale scenes; it does not fully solve fragmented viewpoints capturing disjoint parts of a scene, where both pretrained and fine-tuned models may still fuse disconnected components incorrectly; and important low-level details such as threshold values, file formats, and camera convention are not specified in the supplied text. Another useful distinction is bibliographic: contemporary work on local features continues to use MegaDepth, not MegaDepth-X, as a wide-baseline geometric matching / pose-estimation benchmark, which underscores that MD-X serves a different role in the literature—fine-tuning sparse-view 3D models rather than evaluating pairwise feature pipelines (S. et al., 18 May 2026).

Taken together, MegaDepth-X is best understood as a combined dataset-and-sampling framework for long-tail Internet photo reconstruction. Its novelty lies in the conjunction of cleaned supervision—via MASt3R-SfM, doppelganger pruning, manual validation, and monocular-guided depth filtering—and a sparse-subset sampling distribution that approximates the graph structure of weakly observed Internet scenes. The paper’s empirical claim is not that it solves long-tail reconstruction in a final sense, but that it materially improves robustness to extreme sparsity, repetitive structure, and ambiguous connectivity while retaining much of the performance of pretrained 3D foundation models on more standard benchmarks (Li et al., 24 Apr 2026).