LaVPR: Benchmarking Language and Vision for Place Recognition

Conference: ECCV 2026
Paper: ECCV 2026
Code: https://github.com/oferidan1/LaVPR
Area: Multimodal VLM
Keywords: Visual Place Recognition, Cross-Modal Retrieval, Multimodal Fusion, LoRA, Benchmark Dataset

TL;DR

LaVPR uses Gemini 2.5 Flash to attach 651,865 dense natural-language descriptions to established VPR datasets (GSV-Cities, Pitts30K, MSLS, AmsterTime) and turns both language-vision late fusion and text-to-image "blind" localization into a controlled benchmark on identical visual backbones and original geographic splits: language reliably recovers accuracy on visually degraded subsets (a compact backbone plus one language channel matches much larger vision-only models), while cross-modal localization only becomes viable with LoRA plus Multi-Similarity fine-tuning, lifting R@1 from under 3% zero-shot to 12-38%.

Background & Motivation

Visual Place Recognition (VPR) has spent the past decade inside a tidy paradigm: compress a query image into a global descriptor and retrieve the nearest geo-tagged entry from a database. The descriptor itself kept evolving — NetVLAD's residual clustering, GeM pooling, then MixVPR / SALAD / CricaVPR combining Transformer backbones with attention and optimal-transport aggregation. Yet this purely pixel-driven route is brittle in the real world: heavy rain, long-term seasonal drift, motion blur, and the perceptual aliasing caused by repetitive facades all make even state-of-the-art visual models fail. More fundamentally, a vision-only system simply cannot perform "blind" localization — finding a place from a spoken description alone, which is exactly what emergency response, witness geolocation, and semantic robotics require.

The problem is that mainstream VPR benchmarks are image-only, so there is no text with which to train or evaluate any language-vision model. The few existing attempts each drift off-target: TextPlace treats OCR'd scene text merely as an invariant feature, and while MSSPlace fuses LiDAR, multi-camera imagery, and text — and does provide evidence that multimodal fusion helps — it extends data with MiniGPT-4, has not released its dataset, and its evaluation protocol lacks a controlled ablation (multimodal multi-view inputs are compared against single-view baselines trained on different datasets), so the findings are hard to reproduce. On the text-to-point-cloud side, Text2Pos and KITTI360Pose are mature but address 6-DoF pose regression in a 3D environment from templated descriptions, which is not the VPR paradigm of a single query, a global descriptor, and scalable database retrieval.

So the question "how much does language actually buy VPR?" has never had a clean answer. This paper's premise is that the question is only now answerable: VLMs can generate narrative descriptions of a single image that are far denser than keyword annotations, and open-vocabulary segmentation can ground those descriptions at pixel level — together these turn "retrofit language onto existing datasets" from expensive crowdsourced annotation into an auditable automated pipeline. Core idea: build an auditable pipeline of VLM generation, segmentation grounding, and human review to extend established VPR datasets into a 650K-scale language benchmark, then use strictly controlled comparisons to split the role of language into two measurable questions — how much robustness it adds as an auxiliary channel, and whether it alone can perform blind localization.

Method

LaVPR is not a single model but a combination of a data-construction pipeline and two evaluation tracks: it first builds a text-carrying benchmark, then defines two tasks on top of it — multimodal fusion (V+L) and cross-modal retrieval (L→V) — each with its own controlled protocol.

Overall Architecture

The paper is organized in two halves. The first half constructs the benchmark itself: four existing VPR datasets are extended with dense descriptions, and an audit pipeline of automated screening plus human review guarantees the evaluation sets are trustworthy. The second half runs the two evaluation tracks, which share the same visual backbone and the same original geographic splits, so any performance change can be attributed to a single variable: the presence of language. The fusion track freezes the visual and textual encoders and trains only the late-fusion module, measuring language's gain as an auxiliary channel; the cross-modal track makes text the sole query, measuring how pretrained VLMs behave zero-shot on the "semantic-structural gap" and whether low-rank adaptation can pull them back. The value of this design is that prior work either had only vision, or trained its multimodal and visual components on different datasets, so language's contribution could not be separated from backbone, data, or viewpoint effects.

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
    A["existing VPR datasets<br/>GSV-Cities / Pitts30K / MSLS / AmsterTime"] --> B["description generation and<br/>degraded/scene-text subsets"]
    B --> C["hallucination audit and<br/>human-review loop"]
    C --> D["LaVPR benchmark<br/>651,865 image-text pairs"]
    D --> E["adaptive score fusion and<br/>learned language pooling"]
    D --> F["LoRA + Multi-Similarity<br/>cross-modal alignment"]
    E --> G["multimodal fusion eval V+L<br/>Recall@K"]
    F --> H["cross-modal retrieval eval L to V<br/>Recall@K"]

Key Designs

1. Description generation and degraded/scene-text subset construction: turning every image into a dense, retrievable narrative

Descriptions are generated by Gemini 2.5 Flash with a prompt that explicitly prioritizes permanent architectural features and spatial relations while suppressing transient elements such as pedestrians and vehicles — since queries and reference images are separated by long time spans, only what is stable along the time axis deserves to be a retrieval cue. This is a different regime from sparse keyword annotation: each LaVPR description averages 51.69 words and 23.25 noun entities, the corpus has a unique vocabulary of 209,200, and 28.89% of samples carry scene text such as storefront signage or street signs. The total is 651,865 image-text pairs, split into 529,506 training / 17,608 validation / 104,751 test pairs, all following the original datasets' geographic splits to preserve comparability.

Dataset	Type	Usage	Database images	Query images
GSV-Cities	Urban	Train	529,506	–
Pitts30K-Val	Urban / panorama	Validation	10,000	7,608
Pitts30K-Test	Urban / panorama	Test	10,000	6,816
AmsterTime	Very long-term	Test	1,231	1,231
MSLS-val	Urban / suburban	Test	18,871	740
MSLS-challenge	Long-term	Test	38,770	27,092

What actually makes "the language channel helps" measurable are three deliberately chosen hard subsets: Amstertime-La (31 queries, dominated by scene text), MSLS-Blur (86 queries, blur plus scene text), and MSLS-Weather (86 queries, adverse weather plus scene text). Their shared construction logic is to render the query image's pixel-level information nearly useless; if text still helps there, it is providing an independent semantic cue rather than a redundant copy of the visual features.

2. Hallucination audit and human-review loop: screening with high recall instead of blind cleaning

VLM-generated descriptions hallucinate, and a single hallucination leaking into the benchmark would contaminate every downstream conclusion, so the authors build a three-stage audit pipeline: Phi-3.5-mini-instruct extracts entities from the description, SAM3 grounds them spatially in the image to confirm they exist, and Qwen2-VL-7B-Instruct performs binary verification. The key trade-off is that the automated stage is deliberately tuned for exceptionally high recall (91%) at the cost of precision (22%) — the intent is "better to over-flag than to miss," and the price is that 26% of evaluation samples are sent to human review.

The human review then demonstrates that the raw data is itself high quality: genuine hallucinations affect only about 1% of inspected images, and are typically mild cases where an otherwise precise description contains one non-existent object; these confirmed hallucinations are subsequently removed from the evaluation sets. Manual spot-checks of in-scene text give 93.2% accuracy, with residual errors concentrated in severe occlusions. More importantly, the authors state honestly that all main-paper results are obtained on raw, uncurated training data — the audit pipeline serves as a diagnostic guarantee of evaluation integrity, not as a cleaning step required for optimization. This matters: LaVPR's strong numbers come not from pre-polished data but from the inherently low noise of dense VLM descriptions, plus the method's robustness to minor visual-language noise.

3. Adaptive score fusion and learned language pooling: making language the anchor when vision fails

The fusion track freezes the pretrained visual encoder \(E_v\) and textual encoder \(E_t\) and trains only the fusion module, fully decoupling "what language contributes" from "the backbone was retrained." The authors compare four late-fusion operations: concatenation (CAT, simply \([z_v; z_t]\)), projection and addition (PA, one linear layer per modality then summed), MLP (a shallow network over the concatenation), and their proposed adaptive score fusion (ADS).

ADS differs from all three in that it fuses at the score level rather than the feature level: a small Softmax MLP predicts per-modality weights from the query's visual and textual features, which then weight the similarities. For a query-reference pair \((i,j)\) the two sides' weights are averaged to preserve symmetry, giving a joint similarity of

\[S_{ij} = w_v^{ij}\,S_v^{ij} + w_t^{ij}\,S_t^{ij}, \qquad w_m^{ij} = \left(w_m^{i} + w_m^{j}\right)/2,\ m\in\{v,t\}\]

This form directly addresses the weakness of fixed concatenation. CAT's weights are implicit and static, so when text and visual representations come from independently projected modalities with mismatched numerical scales it cannot dynamically rescale them — consequently CAT actually degrades once paired with learned language pooling (LLP), dropping from 37.0 to 30.2 R@1 on AmsterTime. ADS computes weights per pair, so under blur or bad weather it can shift weight toward the textual side and, combined with LLP, rises to 38.1 — the highest absolute number on that dataset.

As its companion, the LLP module applies self-attention over the hidden states of the textual encoder to compress the whole narrative into a context-aware representation. This differs from taking the [CLS] token or mean-pooling, because in a place description the discriminative content is usually a combination of a few local cues (a particular shop sign, a particular material), which requires attention to select. The authors validate it across all four fusion strategies.

4. LoRA + Multi-Similarity cross-modal alignment: re-orienting a generic semantic space toward place-discriminative features

The cross-modal track's task is precisely defined: given a description \(t_q\), retrieve its corresponding reference image \(I^*\) from a database \(\mathcal{D}\). One easily overlooked detail is that the query image itself never appears in the search database — the text must match a different image of the same geographic place, captured from an entirely different viewpoint and time. This requires aligning fine-grained architectural structure that separates two similar locations, far harder than CLIP-style object-level image-text alignment.

Zero-shot experiments confirm this: CLIP, BLIP, EVA-CLIP-V2, and SigLIP-V2 all sit below 3% R@1 (0.9-2.8% on AmsterTime). The authors call this the "semantic-structural gap" — these models are optimized for object concepts and lack the architectural cues localization needs. Since a new latent space cannot be learned from scratch, the paper instead re-orients the existing shared space: it injects a low-rank update into the pretrained weights, updating only the linear projections,

\[W = W_0 + \Delta W = W_0 + BA, \qquad B\in\mathbb{R}^{d\times r},\ A\in\mathbb{R}^{r\times k},\ r \ll \min(d,k)\]

Ablations show that coverage matters as much as rank: adapting only the Q, K, V matrices underperforms adapting all linear layers including the feed-forward ones, because the structural cues VPR needs are distributed throughout the Transformer blocks; the final configuration adapts all linear layers with \(r=64\). The objective is the Multi-Similarity (MS) loss:

\[\mathcal{L}_{\text{MS}} = \frac{1}{|\mathcal{B}|}\sum_{q\in\mathcal{B}} \left\{ \frac{1}{\alpha}\log\Big[1+\sum_{p\in\mathcal{P}_q} e^{-\alpha(S_{qp}-\lambda)}\Big] + \frac{1}{\beta}\log\Big[1+\sum_{n\in\mathcal{N}_q} e^{\beta(S_{qn}-\lambda)}\Big] \right\}\]

where \(\mathcal{P}_q\) and \(\mathcal{N}_q\) are the positive and negative sets for query \(q\) and \(\alpha\), \(\beta\), \(\lambda\) are hyper-parameters (⚠️ the original equation is corrupted in the PDF extraction; it is restored here from the standard Multi-Similarity loss form — refer to the original paper). Its difference from an ordinary contrastive objective is that it re-weights each pair by similarity and mines hard positives and hard negatives simultaneously — exactly the signal needed to separate geographically adjacent, visually similar places.

Why it must be LoRA together with MS, the ablation answers crisply: full fine-tuning with MS collapses to R@1 ≈ 0.1%, because catastrophic forgetting erases the VLM's broad semantic priors, which conflict directly with VPR's rigid visual-discriminative objective; adding only LLP also gives 0.1%; and using LoRA but reverting to the original contrastive loss still gives 0.1%. Only a low-rank update that bounds the adaptation magnitude while retaining MS's hard-pair mining reaches 12.8% R@1 — neither ingredient works alone.

Loss & Training

Fusion track: visual and textual encoders are fully frozen and every fusion module except CAT (including ADS and LLP) is optimized with the MS loss; ADS feeds the multimodal similarity scores directly into it. Cross-modal track: LoRA covers all linear projections with rank \(r=64\), again optimized with the MS loss; baselines in the comparison use frozen encoders, full fine-tuning, LLP, or the original contrastive loss. Evaluation uniformly uses the retrieval metric Recall@K (R@1/5/10 for the fusion track, R@1/5/10/20 for the cross-modal track), and no training data is mixed across datasets, keeping strict comparability with the vision-only baselines.

Key Experimental Results

Main Results

Table 1 compares the late-fusion mechanisms with MixVPR (512-d) as the visual backbone and BGE-L as the text encoder. ADS + LLP attains the highest 38.1% R@1 on AmsterTime, while on the already-easy Pitts30 the effect of language is nearly saturated (90.8 vs 90.6) — the gain comes mainly from scenes where vision struggles.

Fusion mechanism	AmsterTime R@1	MSLS-val R@1	Pitts30 R@1
Visual-only	35.7	83.2	90.6
ADS (no LLP)	36.1	83.4	90.7
CAT	37.0	81.9	89.9
MLP	31.7	83.2	89.8
PA	31.0	83.0	89.7
ADS + LLP	38.1	83.2	90.8
CAT + LLP	30.2	77.4	85.7
MLP + LLP	32.6	83.4	89.6
PA + LLP	30.2	82.3	89.6

Table 2 makes language the sole query, comparing cross-modal foundation models zero-shot against their LoRA-MS fine-tuned counterparts. All four models improve by an order of magnitude; LoRA-MS-SigLIP-V2 rises from 0.9% to 13.8% on AmsterTime and from 2.6% to 35.7% on MSLS-val.

Model	AmsterTime R@1	MSLS-val R@1	Pitts30 R@1
CLIP	2.0	2.3	10.9
LoRA-MS-CLIP	11.5	38.0	49.3
BLIP	1.3	1.6	8.1
LoRA-MS-BLIP	12.8	38.0	50.2
EVA-CLIP-V2	2.8	3.8	11.4
LoRA-MS-EVA-V2	11.5	32.8	41.5
SigLIP-V2	0.9	2.6	10.1
LoRA-MS-SigLIP-V2	13.8	35.7	49.3

Ablation Study

Table 3 decomposes the training strategy and LoRA configuration on BLIP, and the result is highly convincing: changing any single factor on its own does not work — low-rank adaptation and the Multi-Similarity loss must both be present.

Training strategy	Loss	AmsterTime R@1	R@5	R@10	R@20
Zero-shot (frozen encoders)	–	1.3	4.1	7.0	12.3
LLP only	MS	0.1	0.5	1.0	2.4
Full fine-tuning	MS	0.1	0.4	0.9	1.6
LoRA (QKV, r=16)	Contrastive	4.0	10.1	13.9	19.9
LoRA (QKV, r=16)	MS	8.9	25.0	34.0	44.8
LoRA (QKV, r=64)	MS	10.6	29.1	38.0	49.3
LoRA (all linear, r=64)	Contrastive	0.1	0.4	0.8	1.6
LoRA (all linear, r=64)	MS	12.8	31.4	41.1	53.2

Table 4 views the same family of experiments from another angle: the relative R@1 gain from language is strongly tied to the quality of the visual backbone itself. Supervised convolutional backbones (VGG, ResNet) benefit most, with NetVLAD gaining 199.5% relatively on AmsterTime, whereas self-supervised ViT backbones (DINOv2-L) already carry rich visual-semantic priors, squeezing language's headroom below 7% — and CricaVPR with CAT is even negative on three datasets.

Method	Fusion	Backbone (dim)	Loss	MSLS-B	MSLS-W	Amst.	MSLS-C	Pitts30
La-NetVLAD	CAT	VGG16 (32k)	TRP	+162.3	+136.3	+199.5	+18.3	−1.0
La-CosPlace	CAT	ResNet50 (512)	LMC	+44.9	+29.4	+13.2	+5.7	+0.3
La-EigenPlace	CAT	ResNet50 (2k)	LMC	+43.2	+15.3	+15.5	+3.4	+0.2
La-MixVPR	CAT	ResNet50 (4k)	MS	+21.3	+14.9	+6.0	+0.9	−0.9
La-MixVPR	ADS+LLP	ResNet50 (4k)	MS	+14.8	+10.4	+11.9	+2.2	+0.3
La-SALAD	CAT	DinoV2-L (8k)	MS	+1.2	+3.9	+5.5	−0.9	−0.8
La-CricaVPR	CAT	DinoV2-L (10k)	MS	0.0	+5.2	−9.9	−3.7	−1.5
La-CricaVPR	ADS+LLP	DinoV2-L (10k)	MS	+1.2	+6.6	0.0	+0.6	+0.1

Key Findings

• Language's gain is governed by the strength of the visual backbone, not by the fusion trick. Supervised convolutional backbones leave large gaps in environmental semantics because of classification-style pretraining, and text fills them directly (NetVLAD +199.5%); self-supervised ViTs such as DINOv2 already capture a rich visual-semantic prior, so language's headroom drops into single digits. This also explains why CAT turns negative on CricaVPR while ADS+LLP pulls it back to non-negative.
• Language is a true anchor precisely when vision fails. On the subsets built for degradation, La-MixVPR (ADS+LLP) lifts MSLS-Blur R@1 from 70.9% to 81.4%, MSLS-Weather from 77.9% to 86.0%, and Amstertime-La from 54.8% to 61.3%. The trend is monotonic: the less reliable vision becomes, the larger text's relative contribution.
• On high-dimensional backbones, fixed concatenation loses to adaptive scoring. CAT is unstable on high-dimensional descriptors such as MixVPR (4096-d) and CricaVPR (10752-d), whereas ADS+LLP maintains consistent gains on both, indicating that per-pair dynamic rescaling becomes more necessary as dimensionality grows and modality scale mismatch widens.
• Sequential re-ranking fails catastrophically. Retrieving Top-100 visually and then re-ranking with text yields only 1.2% R@1 on AmsterTime; going text-first then visual gives 4.4%. The bottleneck is that once the first-stage retrieval misses the ground truth, the second modality cannot recover it; joint fusion (La-MixVPR) reaches 37.0%, and 71.6% at R@20. The two modalities must disambiguate each other from the start rather than in relay.
• Horizontal scaling pays off better than vertical scaling. Replacing CricaVPR with a ViT-S backbone (27.2M params, 4.8 GFLOPs) plus a lightweight text encoder, La-Crica-Small reaches 93.0% R@1 on MSLS-Blur with only 6.82 GFLOPs, beating CricaVPR-Base (106.8M params, 17.7 GFLOPs, 91.9%) at roughly 62% less total compute.
• Longer descriptions help the fine-tuned model and break the zero-shot one. As average query length grows from 15 to 60 words, LoRA-MS-BLIP's R@1 rises from 6.5% to 13.2% (a 100% relative gain), whereas zero-shot BLIP peaks at 35 words (3.2%) and collapses back to 1.3%. Generic VLMs are tuned for short, object-centric captions; only a model re-oriented by the MS loss learns to aggregate fine-grained structural cues scattered through a long narrative.
• A hard gap remains between cross-modal and unimodal retrieval. Even the best LoRA-MS-SigLIP-V2 reaches only 13.8% R@1 on AmsterTime versus 35.7% for the vision-only MixVPR; the authors explicitly frame this as an open challenge.

Highlights & Insights

• Turning "does language help?" into an attributable experiment. By freezing the encoders and training only the fusion module, and by placing every variant on the same visual backbone and the same geographic splits, the paper cleanly separates language's contribution from backbone, training data, and viewpoint. This controlled protocol is more valuable than any single performance number, and it directly corrects work like MSSPlace, where "multimodal wins" but it is unclear which factor wins.
• The audit pipeline's trade-off is counter-intuitive but right. The automated screen is deliberately tuned to 91% recall / 22% precision, letting 26% of samples into the human queue. The authors then point out that this very false-positive rate is evidence of high data quality, since genuine hallucinations are only about 1%. Spending human effort on the few true positives, while running the main experiments on raw uncurated data, preserves both evaluation integrity and reproducibility — avoiding the "over-cleaned data makes conclusions unreproducible" trap.
• ADS turns "whom to trust" into a learnable per-pair decision. Rather than assigning text a fixed weight, it predicts weights from the query's image-text features and averages them over the query-reference pair to stay symmetric. The idea transfers to any primary-plus-auxiliary modality retrieval or fusion setting (LiDAR + camera, audio + video), especially when the auxiliary modality fails entirely on some samples.
• The LoRA-plus-MS failure controls are unusually informative. A single ablation family rules out three shortcuts at once — full fine-tuning, swapping only the loss, adding only pooling — and all of them fail in the same way, collapsing to R@1 ≈ 0.1%. For any work adapting a strongly pretrained model to a new domain, this is a reusable warning: adaptation capacity and objective must be co-designed, otherwise the model does not merely fail to learn, it destroys what it already knew.

Limitations & Future Work

• Gains on strong self-supervised backbones are limited or even negative. The authors acknowledge that when the underlying visual features are already highly discriminative, injected linguistic context can occasionally introduce sub-optimal priors (see La-CricaVPR + CAT at −9.9% on AmsterTime in Table 4). The benchmark's value therefore lies more on the weak-backbone and degraded-scene side.
• Cross-modal blind localization is far from usable. The best cross-modal baseline reaches only 13.8% R@1 on AmsterTime against 35.7% unimodal; the authors raise this as a community challenge rather than packaging it as a solved result.
• Only late fusion is systematically evaluated. The authors note that preliminary mid-fusion experiments degraded performance, but the trade-off between early fusion's tight semantic alignment and the preservation of fine-grained discriminative visual features is not systematically explored — it lies outside this work's scope.
• Description generation relies on a single VLM and an offline pipeline. Every description comes from Gemini 2.5 Flash, so the benchmark inherits that model's descriptive bias (favoring architecture and spatial relations, down-weighting pedestrians). Whether a different generator would change the conclusions is not cross-validated. Moreover, being generated offline, the descriptions do not reflect the colloquial, incomplete, even emotional distributions of real human requests — precisely the most realistic inputs for blind localization.
• Improvement directions: extend the query side to genuine human spoken descriptions (a human-rewritten subset could serve as a stress test); introduce stronger alignment objectives for the cross-modal track (for example hierarchical mining of geographically adjacent negatives) rather than relying on the MS loss alone; and systematically quantify the early/mid/late fusion trade-off.

Related Work & Insights

• vs MSSPlace: MSSPlace likewise fuses multiple modalities with text (extending Oxford RobotCar and NCLT via MiniGPT-4) and provides evidence that fusion helps, but its evaluation protocol lacks a controlled ablation — multimodal multi-view inputs are compared against single-view baselines trained on different datasets — and neither the dataset nor a detailed construction protocol has been released, making the findings hard to reproduce. LaVPR differs through its strictly controlled setup and public audit pipeline, which isolate language's contribution.
• vs TextPlace: TextPlace treats scene text extracted by OCR as an invariant feature, depending on the natural presence of strongly discriminative cues such as street signs and shop names. LaVPR is not limited to scene text: language carries the semantics of the whole scene (materials, structures, spatial relations), with scene text only one subset (about 29%), so it also works where no signage exists.
• vs Text2Pos / KITTI360Pose: Text2Pos-style text-to-point-cloud localization performs coarse-to-fine 6-DoF pose regression in a 3D environment, relying on templated descriptions of spatial relations. LaVPR's cross-modal task stays within the VPR paradigm — a single natural-language query retrieving a global descriptor from a scalable database — and is built directly on established VPR train/eval benchmarks, so its results are directly comparable to vision-only ones.
• vs vision-only VPR (MixVPR / SALAD / CricaVPR): These methods spend all their capacity on descriptor aggregation and training objectives and are already strong in non-degraded scenes (90%+ on Pitts30), but they degrade under blur, weather, and very long-term change. LaVPR leaves their backbones untouched and adds only one language channel for consistent gains, with far larger benefits for small backbones — suggesting that multimodal horizontal scaling may be a better deal than stacking more visual Transformers under constrained compute.

Rating

• Novelty: ⭐⭐⭐⭐ Not a new model, but it converts "is language useful for VPR?" from a slogan into a controlled, measurable benchmark, and open-sources 650K image-text pairs plus the audit protocol, filling an infrastructure gap.
• Experimental Thoroughness: ⭐⭐⭐⭐ Covers 6 datasets, 3 degraded subsets, 5 visual backbones, 4 text encoders, 4 fusion mechanisms, and a complete LoRA/MS ablation, and reports negative gains explicitly; it loses a point for not systematically evaluating early/mid fusion and for not cross-validating the description generator.
• Writing Quality: ⭐⭐⭐⭐ The two evaluation tracks are organized clearly, the data-construction and audit procedures are well explained, and negative results are not glossed over; some equations are garbled in the typeset PDF (restored here from standard forms).
• Value: ⭐⭐⭐⭐ The benchmark and code are open and the controlled protocol is directly reusable by follow-up work; the cross-modal track's 13.8% vs 35.7% gap is a well-defined and practically meaningful open problem.

Related Papers

• [CVPR 2026] DialogueVPR: Towards Conversational Visual Place Recognition
• [ECCV 2026] LongVQUBench: Benchmarking Long-Term Video Quality Understanding of Vision-Language Models
• [ECCV 2026] Benchmarking MLLMs on Mistake Recognition and Explanation in Single-Step Components of Cooking
• [ECCV 2026] Divide and Align: Disentangled Vision-Language Learning for Text-Based Person Anomaly Search
• [ECCV 2026] Read or Ignore? A Unified Benchmark for Typographic-Attack Robustness and Text Recognition in Vision-Language Models