Burst-Matched Alignment for Place Recognition on Cheap City Meshes

A place-recognition system identifies where a photo was taken by retrieving the most similar image from a geo-tagged reference database. Building that database has traditionally meant driving camera cars through a city. MeshVPR showed a cheaper option: render the database from openly available 3D textured city meshes. The cost is a domain gap. Standard models fall from 96.3% to 76.9% Recall@1 in San Francisco when the database is synthetic. Teacher-student feature alignment recovers only part of this, and a high-quality mesh adds 15.7 Recall@1 points over a low-quality one. So the remaining lever is to pay more for rendering and mesh acquisition. This note argues that much of what goes wrong in cheap renders has a specific, measurable shape. Rendering changes visual burstiness, the tendency of the same local pattern to repeat many times within one image. Low-resolution texture maps get tiled across walls, and foliage and ornament collapse into a few coarse, self-similar patches. Retrieval research has long known that such repeated patterns swamp the rare, distinctive details a match depends on.

The proposal is to train the place-recognition model so that its count of repeated patterns on a render matches its count on the corresponding real photo. The model then learns to discount repetition that the renderer introduced, rather than repetition that exists in the world. If this works, a low-quality mesh with burst-matched alignment should recover a large share of the high-quality-mesh advantage. That would turn an algorithmic change into a direct reduction of synthetic-data overhead. The first thing to test is diagnostic: does per-query mismatch in burstiness between render and photo actually predict which queries fail?

Core Claim

The synthetic-to-real gap in mesh-based place recognition is concentrated in a mismatch of intra-image repetition statistics. Aligning those statistics explicitly is more sample- and cost-efficient than aligning whole descriptors alone.

Concretely, let $f_i$ be the local patch features a backbone extracts from an image. Following the "Soft Count" of VLAD-BuFF, let

$$c_i = \sum_{j} \sigma\big(\alpha \cos(f_i, f_j) - \beta\big)$$

be the soft number of near-duplicates of patch $i$ in the same image. Here $\sigma$ is the logistic function and $\alpha, \beta$ are learned. VLAD-BuFF uses $c_i$ to down-weight bursty patches before pooling. We add one term to MeshVPR's alignment objective. Given a render $r$ and its corresponding real photo $p$, we match the student's counts on the render to the frozen teacher's counts on the photo:

$$\mathcal{L} = \lVert g_s(r) - g_t(p) \rVert^2 + \lambda \, D\big(\{\log c_i^{s}(r)\}, \{\log c_i^{t}(p)\}\big)$$

In this objective, $g_s, g_t$ are student and teacher global descriptors, and the first term is MeshVPR's MSE. $D$ is a patch-wise squared difference where spatial correspondence is reliable, and a one-dimensional Wasserstein distance between count histograms otherwise. $\lambda$ weights the new term. The descriptor MSE tells the student what the output should be. The count term tells it which patches to ignore on the way there. That second signal is exactly what differs between a tiled texture and a real facade.

Research Program

The work has three parts, ordered so that the first can kill the idea cheaply.

First, a measurement study on MeshVPR's paired data in San Francisco and Berlin, at both mesh qualities. We compute Soft Count distributions with a frozen DINOv2 backbone on renders and their real counterparts. We test two things. Renders should have systematically heavier-tailed count distributions than photos, and more so for the low-quality mesh. The per-location mismatch in counts should predict per-query retrieval failure under MeshVPR's aligned model, after controlling for simpler image statistics such as sharpness and color histogram distance.

Second, the method itself. We use a DINOv2 student with a VLAD-BuFF-style burst-aware aggregator, trained with the objective above. This part draws its roles from prior work in three ways. The count operator and its 64-dimensional projection come from VLAD-BuFF, which applies them only to real imagery. The teacher-student recipe comes from MeshVPR, which never inspects the repetition statistics of renders. The motivation that repetition corrupts pooled similarity comes from On the burstiness of visual elements, which fixes it with fixed sub-linear weighting at retrieval time.

We also take one constraint seriously. Visual Place Recognition with Repetitive Structures argues that real repetition, such as a building's own window grid, is itself a distinctive signature. It gains 11.97% recall on Pittsburgh by grouping repeated elements rather than deleting them. Uniform suppression would therefore be wrong. This is why the target of the count term is the teacher's counts on the real photo: repetition that is real is preserved, and only render-induced excess is penalized. It is also why we do not simply suppress everything self-similar.

Third, the overhead question. We express results as Recall@1 against rendering and storage cost for low- versus high-quality meshes. The target is to see how much of the 15.7-point quality premium burst-matched alignment buys back.

Baselines

The naive reference point is an off-the-shelf model with no alignment queried against the rendered database. For this setting MeshVPR reports 76.9% Recall@1 in San Francisco, against 96.3% on a real database.

The strongest existing method is MeshVPR's MSE teacher-student alignment on SALAD. It reports Recall@1 of 88.8% in San Francisco (up from 76.9%) and 82.0% in Berlin (up from 69.8%). We rerun it ourselves with the same DINOv2 backbone, the same paired training data, the same number of epochs, and the same hyperparameter search budget as our method, so that differences come only from the objective. For Paris and Melbourne, no reported result was found on Lacuna for MeshVPR, so our reruns will be the reference there. Beating this baseline means a gain over MSE alignment on the same mesh that holds across at least San Francisco and Berlin. It also means, for the overhead claim, that our low-quality-mesh model lands clearly closer to the high-quality-mesh MSE baseline than the low-quality-mesh MSE baseline does.

The critical ablation isolates the count-matching term. It swaps SALAD for a VLAD-BuFF aggregator inside MeshVPR's MSE alignment, with $\lambda = 0$. No reported result was found on Lacuna for VLAD-BuFF on mesh-rendered databases, so this must be run. It separates two effects: the benefit of a burst-aware architecture, and the benefit of supervising burstiness toward real-photo statistics. A second ablation replaces the teacher-matched target with uniform suppression, pushing all counts down. This tests the repttile concern directly.

Why not just…?

Why not just drop VLAD-BuFF into MeshVPR? This is the most serious threat, and the proposal is built so the experiment settles it. VLAD-BuFF learns to discount self-similar patches from real street-view data, so it may already discount tiled render textures without any extra signal. It may not, because a render's tiled texture and a real repeated window grid look alike to a self-similarity score. Only a paired target can distinguish them. If the $\lambda = 0$ ablation matches the full method, we will report that the aggregator alone suffices. The contribution then reduces to the measurement study and the cost analysis, and we would say so.

Why not just buy the high-quality mesh? MeshVPR shows that it helps by 15.7 Recall@1 points but does not close the gap; even that setup falls well short of a real database. It also brings heavier storage, rendering time, and manual repair around vegetation and tunnels. Our method is compatible with high-quality meshes, and we run it on both. The claim under test is a trade-off: how much Recall@1 per unit of rendering cost. It is not a claim that cheap meshes beat expensive ones.

Why not just augment the renders or train alignment longer? Generic photometric augmentation, blur, and longer MSE training are cheap, and they may narrow the gap. None of them targets the tiled-texture artifact specifically. We include an augmentation-plus-extended-training MSE baseline at matched compute. If it equals burst matching, the burstiness diagnosis is not what drives the gain.

Proposed Experiments

The decisive first check is the correlation study. Among San Francisco and Berlin queries that MeshVPR's aligned model gets wrong, the render-versus-photo count mismatch at the retrieved and true locations should be significantly larger than for correct queries. It should also remain predictive after we regress out sharpness and color statistics. If it does not, we stop and report a negative result.

If the check passes, we train all methods on MeshVPR's paired data for both mesh qualities and evaluate Recall@1 and Recall@5 on San Francisco, Berlin, Paris, and Melbourne. We plot each model against rendering time and mesh storage.

Two further checks guard against overfitting to renders. The first is transfer: a model trained with burst matching should not lose accuracy on real-database benchmarks such as Pitts250k, Tokyo 24/7, and Nordland, compared with the same model trained without it. The second is qualitative: we visualize per-patch weights on renders, expecting tiled wall textures to be down-weighted while real window grids keep weight similar to the photo. A reduction in the low-versus-high-quality mesh gap under burst matching, beyond what the VLAD-BuFF ablation and the augmentation baseline achieve at equal compute, is the result that would justify the claim.