Can 1,000 open scenes make a state-of-the-art 3D model better?

01The question

If a small open sample already moves a frontier model, what is the full dataset worth?

This note is a dataset assessment before it is a modeling exercise. OverMaps-1k is the open sample of the OVER dataset: about a thousand outdoor and indoor scenes reconstructed with COLMAP, with metric camera poses from phone AR tracking — released on Hugging Face so anyone can use it for research and reproduce every number in this note. R3 is a relative-pose regression model built on a Depth Anything 3 / DINOv2 backbone, trained on a large mix of public datasets that does not include ours. We wanted to know whether further training on the open sample makes R3 better, and at what cost — because the same playbook applies to any foundation model you want to specialize on new data, and because the sample is a deliberately small probe: the corpus behind it is about 270× larger. What a thousand scenes do to a frontier model is the conservative floor of what the dataset can offer.

Three numbers hold the story together; the rest of the note reconstructs them one at a time.

02The gain

On our domain, pose accuracy improves — visibly

Our first study used the «short» variant (windows of up to 32 views). We evaluate on 50 scenes held out of training entirely, measuring trajectory error (ATE) normalized by window extent. The base model starts at 37.7%; after fine-tuning we land around 32%, with rotation error dropping from 5.0° to 3.7°. Three independent runs converge to the same numbers, so this is not one lucky checkpoint.

The most direct way to see it is to overlay the camera trajectory the model reconstructs onto ground truth. Where the base model drifts, the adapted one stays on the path.

The same thing, explorable in 3D. What R3 outputs — and what fine-tuning improves — are camera poses: below you can inspect them directly. The reconstructed trajectory (ground truth, base, fine-tuned, Sim3-aligned) with a sparse context cloud. Drag to rotate, scroll to zoom; the buttons toggle each trajectory.

One clarification on method: we show poses, not a dense fused surface. R3's per-frame depth is good, but its scale is not perfectly consistent across frames, and fusing all depths into a single surface fragments on our outdoor scenes. Pose is what we measure, what fine-tuning improves, and exactly what this view shows.

Online reconstruction, frame by frame

In the style of R3's own demo: on the left the input stream and the rendered flythrough — the point cloud accumulating frame by frame, filmed by an external camera travelling above the reconstruction (rendered offline, as R3's project page does; predicted depth as an inset top-right, trajectory colored blue→red over time); on the right the reference point cloud, the full COLMAP reconstruction of the area — static, mouse-explorable, with the trajectory and current position. We compare our fine-tuned model against the R3 base; below, cumulative position error grows in real time.

03A risk we had to defuse

Pose-only fine-tuning tends to collapse depth

R3 does not only predict pose: from the same shared backbone it also estimates a per-frame depth map. This is where our first attempt went wrong. A pose-oriented fine-tune, at a learning rate high enough to actually move the pose metrics, drags the backbone with it — and the depth head, reading those same features, came out flattened. Depth spread had collapsed roughly four-fold, and point clouds were flat and unusable.

The fix that held is structural, not a learning-rate compromise: during training we distill depth from a frozen metric DA3 teacher and keep the depth head frozen [1,8]. The teacher acts as an anchor, forcing the shared features to stay compatible with the original depth while pose specializes. Depth structure stays glued to the base at every checkpoint.

04The cost

Features drift, and external benchmarks feel it

The most insidious cost is forgetting. As we improve on our domain, the shared representations move away from the distributions R3 was trained on, and performance there degrades. We measure it on TUM-RGBD [7], a classic odometry benchmark with motion-capture ground truth — close in flavor to R3's training data, and therefore a good proxy for how much the model has forgotten its home world.

A naive fine-tune degrades TUM pose error by 2.48× on average. It is not a bug and not an artifact: as we show below, the degradation is perfectly gradual as you move away from the base weights. So the question becomes operational — can we keep the in-domain gain without paying this price?

05Countermeasures

Four families of remedies, one measuring stick

The catastrophic-forgetting literature offers four broad families — parameter-efficient methods, replay of original-style data, functional distillation, and weight merging [8,9,10,11,12] — and we went through them one by one, on the same protocol so they can be compared.

Two of these deserve a remark. LoRA was the «obvious» candidate: freezing the base weights and learning small adapters should forget less; in practice it did worst of all, a sign that forgetting here depends more on how aggressively you update than on how many parameters you touch. LwF was the most elegant idea — anchoring the model to its own original behavior without any ground truth — and it delivered the best in-domain rotation of the study, but on forgetting it did not beat plain replay.

No single training-time lever cancels forgetting, but the levers stack. Replay — and remarkably little of it; a dozen rehearsal scenes anchor as well as a hundred — brings forgetting from 2.48× down to 1.52×. Weight interpolation applied on top brings it to 1.15×.

06The winning lever

WiSE: one knob on the trade-off, at almost zero cost

The strategy that moves the needle most is also the cheapest. WiSE-FT [10] does not touch training at all: it takes the already-adapted model and linearly interpolates its weights with the base. A single coefficient α decides where to stop — α=0 is the base, α=1 the full fine-tune — and every value in between is a different trade-off. It costs a minute of CPU and hands you the entire frontier between gain and forgetting.

The shape of the curve — a smooth ramp, not a cliff — tells us something about the problem. If R3 had somehow «memorized» its benchmarks, the smallest step away from the base weights would crater them; instead they degrade gradually and predictably. That is ordinary representation drift, not benchmark overfitting — and precisely because of that, it is controllable.

07From study to deployment

Repeating everything on the long model — and beating the public checkpoint on its own turf

Everything so far used the short variant. But the checkpoint people actually deploy for video is r3_long (32–100 view windows, the default of R3's long/streaming modes). So we re-ran the whole recipe there: same data, same protections, initialized from r3_long, evaluated at 100 views — and derived WiSE interpolations against r3_long itself, so that every point on the dial is a drop-in replacement for the deployed model.

In-domain, our models cut the deployed checkpoint's error by a quarter (34.6→26.0 mean, 27.5→21.1 median) and even beat the stronger short baseline. On TUM the α=0.3 interpolation gives back only 6% — and stays better than the public short model.

Then we stress-tested where fine-tunes usually embarrass themselves: a fast 800-frame FPV video, completely unrelated to our data, reconstructed in R3's streaming mode. To score it honestly we built an independent reference with COLMAP (all 800 frames registered, 0.66 px mean reprojection error) and measured every model's trajectory against it after Sim3 alignment.

Two honest footnotes. First, α still matters out of domain: the α=0.5 point, best in-domain, pays a real price on this video (5.59% mean error) — the dial is a dial, not magic; we recommend α=0.3 as the default operating point and α=0.5 when the workload matches our domain. Second, our first WiSE merges for the long model were built against the wrong base (the fine-tune's init was the short model), and their out-of-domain behavior collapsed — re-running the fine-tune initialized from r3_long fixed it. Weight interpolation only works along the actual training trajectory; the base you interpolate with must be the base you trained from.

08How much data does it take?

The scaling picture

In the initial 32-view study, most of the measured gain arrived with the first ~250 scenes, with a slower improvement up to 950 — and per-scene analysis shows the lift is broad rather than scene-specific: the fine-tune shifts a general prior (wide-baseline calibration on real-world captures) rather than memorizing places. We are re-measuring the curve in the deployment (100-view) regime; this section will be updated with those results.

Where the remaining error lives is just as informative: auditing all 50 held-out scenes, the reconstructions our models still fail on are the classic hard cases — dense vegetation, large symmetric structures, repetitive interiors — and captures whose photo ordering hops around the scene, which is structurally hostile to any online, causally-attending model. Fixing that tail is not a matter of more of the same data; it points at targeted data diversity and at sequence-aware training windows, both of which are actionable.

09Why our data matters

The gap is one of performance, not appearance

We were tempted to tell ourselves that our data is visually alien to R3's world. We checked, by measuring domain distances in the backbone's feature space, and the verdict is that the appearance distance is modest: our data, ARKitScenes and TUM form nearly an equilateral triangle, with TUM only 1.26× closer to R3's training mix than our scenes are. The 2-D projection suggested a cleaner separation than actually exists. That is not the argument.

The solid argument is about the task: base R3 is roughly ten times worse on our real-world scenes than on its benchmarks. Reported performance reflects the training distribution; taken to real captures, the model is much weaker. That performance gap is exactly the space our data fills.

10Conclusions

What a thousand open scenes proved — and what the corpus could

Fine-tuning on our data works, reproducibly, on both released variants of the model — and on the deployment checkpoint it is not a trade-off but a strict upgrade at the right operating point: a quarter less trajectory error on our domain, a measured edge even on out-of-domain footage, and external benchmarks within a few percent of the base. Depth is preserved by construction, distilled from a frozen teacher. Forgetting is real but governable in steps: a little replay during training, weight interpolation after it, and — crucially — interpolating against the exact base the fine-tune started from.

The practical consequence: the value of a dataset is measured where the gap is real — on real-world performance. One thousand open scenes — the sample anyone can download today — moved a state-of-the-art model by a quarter on that axis, at essentially zero cost elsewhere, using one GPU and a recipe that transfers to any foundation model with the same shape.

That is the floor this study establishes, and also its limit. Our audit of the residual errors says the next gains will come not from more of the same scenes but from targeted diversity — more geographies, more capture styles, more hard cases. Diversity at scale is what the full OVER corpus holds: roughly 270× the scenes of the open sample, indoor and outdoor, built with the same metric-pose recipe throughout. Assessing what that does to frontier models — for pose, for depth, for tasks this note never touched — is research we want to see happen, and the open sample is the standing invitation. If your experiments outgrow the sample, so does the dataset: reach out.

—References

1. R3 — 3D Reconstruction via Relative Regression. The base model under study (relative-pose regression, DA3 backbone).
2. Depth Anything 3. Backbone and frozen metric-depth teacher used for distillation.
3. Oquab et al. DINOv2: Learning Robust Visual Features without Supervision. 2023.
4. Schönberger & Frahm. Structure-from-Motion Revisited (COLMAP). CVPR 2016.
5. Apple. ARKit — visual-inertial odometry providing the metric reference poses of our scenes.
6. Baruch et al. ARKitScenes. NeurIPS Datasets & Benchmarks 2021 — indoor replay anchor.
7. Sturm et al. A Benchmark for RGB-D SLAM Evaluation (TUM-RGBD). IROS 2012.
8. Li & Hoiem. Learning without Forgetting. ECCV 2016 / TPAMI 2017 — functional distillation; conceptual basis of our pose self-distillation.
9. Hu et al. LoRA: Low-Rank Adaptation of Large Language Models. ICLR 2022.
10. Wortsman et al. Robust Fine-Tuning of Zero-Shot Models (WiSE-FT). CVPR 2022.
11. Ibrahim et al. Simple and Scalable Strategies to Continually Pre-train LLMs. 2024 — small-fraction replay suffices to mitigate forgetting.
12. Yadav et al. TIES-Merging / Wortsman et al. Model Soups — context on weight merging.
13. Umeyama. Least-squares estimation of transformation parameters. TPAMI 1991 — alignment used for ATE (rotation from orientations, scale/translation from centers).