EPRBench: Benchmark for Event-Based VPR

• EPRBench is a benchmark dataset and evaluation suite for event-based visual place recognition that incorporates large-scale event streams and scene-level semantic annotations.
• It combines high-definition event data from Prophesee EVK4-HD, curated textual descriptions from LLMs, and a standardized Recall@N protocol across diverse urban scenes.
• The introduced SG-VPR baseline fuses visual and language features with spatial attention, achieving up to 97.1 R@10 and providing interpretable semantic reasoning.

EPRBench is a benchmark dataset and evaluation suite for event stream-based visual place recognition (VPR), introduced to address the scarcity of dedicated event-VPR resources and the limited use of semantic information in event-based retrieval pipelines (Wang et al., 13 Feb 2026). It is presented as a high-quality, high-definition benchmark that supports both robust event-based place recognition and semantic-aware, language-integrated VPR. The benchmark combines large-scale event data, scene-level textual annotations generated by LLMs and refined through human annotation, a standardized evaluation protocol based on retrieval Recall@N, and a multimodal baseline called SG-VPR. In the paper’s own descriptions, EPRBench is said to comprise “10K event sequences and 65K event frames,” while the dataset section reports 13,109 event scene samples; each sample consists of 5 event frames, so both descriptions indicate a substantially larger resource than earlier event-VPR datasets (Wang et al., 13 Feb 2026).

1. Motivation and problem setting

EPRBench was motivated by two bottlenecks in event-based VPR research. The first is dataset scarcity: the paper states that there have been very few public event-VPR datasets, and that existing resources are small in scale and limited in diversity (Wang et al., 13 Feb 2026). The second is the lack of semantics in most existing methods. Prior work in this area is described as focusing primarily on feature learning from raw visual inputs, without exploiting scene descriptions or linguistic cues that might help disambiguate visually similar places.

The benchmark is grounded in the sensing properties of event cameras. These sensors capture brightness changes asynchronously and are described as offering high dynamic range, resilience to low light and glare, reduced motion blur under high-speed motion, low latency, and low power consumption. Those properties make event cameras attractive for VPR under conditions that are challenging for conventional visible-light cameras, including low illumination, overexposure, and high-speed motion.

Within that context, EPRBench is designed not merely as a repository of event streams, but as an integrated benchmark for evaluating robustness to viewpoint variation, weather variation, illumination change, event sparsity, and noise. The inclusion of semantic annotations also positions it as infrastructure for semantic-aware and language-integrated VPR, rather than only appearance-based retrieval.

2. Dataset composition, sensing hardware, and acquisition regimes

EPRBench was collected with a Prophesee EVK4-HD event camera at a resolution of $1280 \times 720$, which the paper explicitly contrasts with earlier event datasets based on lower-resolution sensors such as DAVIS 346 (Wang et al., 13 Feb 2026). The dataset is organized around more than 1,000 urban scenes and is divided into three scene categories intended to cover most typical city environments: Campus, Park, and Road.

Category	Scenes	Samples
Campus	315	2,766
Park	351	5,176
Road	356	5,167

The total reported split is random and follows a $7{:}1{:}2$ ratio, with 9,220 training samples, 1,273 validation samples, and 2,616 testing samples. Each event sample consists of 5 event frames. A possible source of confusion is the paper’s dual scale description: the abstract and introduction refer to “10K+ event sequences,” whereas the dataset section reports 13,109 event scene samples. The paper presents both counts without resolving the terminology; a plausible interpretation is that “event sequences” and “event scene samples” are used at slightly different granularities.

Data collection used two complementary acquisition modes. A vehicle-mounted setup was used to capture road scenes, wide field-of-view urban environments, and rich semantic context over larger spatial extents. A handheld or manual setup was used to capture finer-grained details, smaller-scale environments, and viewpoints that vehicle-mounted recording could miss. The benchmark further includes multiple viewpoints for each scene, various weather conditions, and different lighting conditions. The paper’s comparison table identifies clear sky, rainy day, daytime, and night time as represented conditions.

This acquisition design matters because VPR performance is strongly shaped by nuisance factors such as viewpoint displacement and appearance change. EPRBench was constructed so that such variations are not incidental artifacts of collection, but deliberate parts of the benchmarked problem.

3. Semantic annotation pipeline and language supervision

A distinctive feature of EPRBench is the inclusion of scene descriptions intended to support language-integrated retrieval and interpretable reasoning (Wang et al., 13 Feb 2026). The paper argues that this is especially useful for event streams because event data are sparse and difficult to interpret directly, while semantic cues such as landmarks, scene layout, and structural clues can improve place disambiguation.

The annotation pipeline is described as a three-stage procedure. In the first stage, the authors construct 50,000 image sample pairs from the original event data, with positive and negative pairs balanced. Using a template prompting strategy, they ask Doubao Pro-1.6 to generate Chinese chain-of-thought-style scene judgments or descriptions, yielding an initial Chinese CoT dataset.

The second stage performs cross-lingual translation and refinement. The Chinese CoT data are translated into English using NLLB-200-3B. The translated text is then further cleaned and corrected with DeepSeek using targeted prompts, specifically to address expression deviation and grammar errors. The result is a higher-quality English CoT dataset.

In the third stage, Qwen2.5-VL-7B is fine-tuned using LoRA and supervised fine-tuning so that it becomes a “scene expert” capable of generating semantic descriptions for the dataset images. The abstract adds that these LLM-generated scene descriptions are subsequently refined through human annotation. The paper does not provide a detailed manual-annotation protocol, but it clearly assigns human refinement a corrective role in ensuring semantic accuracy.

This semantic layer changes the character of the benchmark. EPRBench is not restricted to feature extraction from event frames; it also supports training and evaluation of systems that fuse event-derived representations with linguistic scene descriptions, and it enables reasoning-style outputs beyond a scalar similarity score.

4. Evaluation protocol and baseline benchmarking

The evaluation metric used in EPRBench is Recall@N with $N \in \{1,5,10\}$ (Wang et al., 13 Feb 2026). A retrieval is counted as correct if at least one of the top-$N$ retrieved database images has the same location label as the query. This is the conventional retrieval-based formulation of VPR, and the benchmark adopts it in a systematic way across a broad set of baselines.

The paper reports results for 15 representative or state-of-the-art VPR methods. These include CNN-based global descriptors such as CosPlace, MixVPR, and EigenPlaces; Transformer-based local aggregators such as CricaVPR, BoQ, and SALAD; re-ranking or two-stage methods such as R2Former, SelaVPR, and Pair-VPR; and additional methods including Deep visual geo-localization, EMVP, SuperVLAD, ImAge, PRGS, and FOL.

On EPRBench itself, the best reported result is SG-VPR with $94.3$ R@1, $96.1$ R@5, and $97.1$ R@10. Other strong baselines reported in the same comparison include SuperVLAD with $93.3$ R@1, CricaVPR with $92.6$ R@1, Pair-VPR with $91.9$ R@1, and FOL with $7{:}1{:}2$0 R@1. The paper therefore presents EPRBench as challenging but not saturated: modern methods can reach high retrieval accuracy, yet the benchmark still exposes differences among top-performing systems.

The authors also report transfer-style evaluations on NYC-Event-VPR. On the event modality, SG-VPR reaches $7{:}1{:}2$1 R@1, compared with $7{:}1{:}2$2 for CricaVPR and $7{:}1{:}2$3 for MixVPR. On the RGB split, SG-VPR achieves $7{:}1{:}2$4 R@1. These comparisons are used to argue that semantic guidance is particularly helpful on sparse and noisy event data under severe urban viewpoint change.

5. SG-VPR: semantic-guided multimodal fusion

SG-VPR is the paper’s proposed multimodal baseline for EPRBench, and it operationalizes the benchmark’s central claim that scene semantics can improve event-based VPR (Wang et al., 13 Feb 2026). The method turns event-based place recognition into a vision-language fusion problem.

The raw event stream is written as

$7{:}1{:}2$5

Because the event stream is asynchronous, it is converted into a synchronous 2D frame by accumulating events over a time window $7{:}1{:}2$6, producing

$7{:}1{:}2$7

The event frame is encoded by DINOv2 to obtain patch embeddings

$7{:}1{:}2$8

and a global class token

$7{:}1{:}2$9

In parallel, a text description $N \in \{1,5,10\}$0 is processed by a frozen CLIP text encoder to yield word-level features

$N \in \{1,5,10\}$1

and a sentence-level feature

$N \in \{1,5,10\}$2

Global cross-modal fusion concatenates the global visual and textual tokens and projects them through an MLP: $N \in \{1,5,10\}$3 This global vector is intended to capture holistic scene context. Local fusion then uses the word-level text tokens to score visual patches by maximum cosine similarity: $N \in \{1,5,10\}$4 The method retains the top-$N \in \{1,5,10\}$5 fraction of patches,

$N \in \{1,5,10\}$6

with $N \in \{1,5,10\}$7 in the reported experiments. For selected patches $N \in \{1,5,10\}$8, semantic injection is defined as

$N \in \{1,5,10\}$9

This “spatially attentive token selection” mechanism is the core local alignment module: text is used to identify and enhance semantically meaningful visual regions, often corresponding to landmarks or stable structural elements.

After local enhancement, the patch tokens are reshaped into a 2D feature map $N$0. The final representation uses a modified spatial pyramid in which the global semantic vector acts as an anchor and GeM pooling is applied at $N$1 and $N$2 scales: $N$3 The result combines holistic semantics, mid-level geometry, and finer spatial detail within a single descriptor.

6. Training objective, interpretability, and benchmark significance

SG-VPR is trained with a combined objective consisting of Multi-Similarity metric learning and symmetric InfoNCE cross-modal alignment (Wang et al., 13 Feb 2026). The Multi-Similarity term is written as

$N$4

with $N$5, $N$6, and $N$7. The symmetric InfoNCE term is

$N$8

with $N$9. The total loss is

$94.3$0

where $94.3$1.

The paper presents SG-VPR as interpretable for three stated reasons: retrieval is explicitly grounded in generated scene descriptions; the patch-selection mechanism is based on text-visual similarity; and an auxiliary LLM decoder can translate the fused representation back into natural-language reasoning. This decoder is intended to produce chain-of-thought-style explanations describing shared landmarks, scene structure, semantic consistency, or mismatched geometry in failure cases.

The benchmark significance of EPRBench lies in this coupling of dataset design, semantic supervision, and systematic evaluation. It provides a large-scale event-based VPR dataset, scene descriptions generated by LLMs and refined by human annotation, benchmark results for 15 VPR methods, and a semantic-guided baseline that integrates event and text features. At the same time, the paper acknowledges a limitation: the “scene expert” used for text generation is fine-tuned to the dataset distribution, which limits transferability to open-world RGB scenarios. This suggests that one future direction is the construction of a more general-purpose scene expert, while another is the use of EPRBench as a testbed for broader semantic-guided event perception problems beyond place recognition.