SynVS-Day: Synthetic Daytime Vehicle Search Dataset

• SynVS-Day is a synthetic multi-camera dataset curated from virtual cities, enabling joint vehicle detection and re-identification specifically in daytime scenes.
• It minimizes environmental confounds by isolating day scenes, ensuring controlled evaluations with exact, simulation-derived annotations.
• Designed for end-to-end vehicle search, the dataset supports benchmark metrics like mAP and CMC Top-1, helping improve unified detection and re-ID models.

SynVS-Day is a synthetic, multi-camera vehicle search dataset introduced in the CLIPVehicle benchmark for end-to-end vehicle search, that is, joint detection and re-identification directly on raw surveillance frames rather than on pre-cropped vehicle patches. It is curated from the Synthehicle multi-vehicle, multi-camera tracking corpus of virtual cities, restricted to daytime scenes, and positioned alongside the real-world CityFlowVS dataset and the mixed-condition synthetic SynVS-All dataset to support both controlled development and fair evaluation of vehicle search systems (Wang et al., 6 Aug 2025).

1. Benchmark role and problem setting

SynVS-Day was introduced to address a dataset gap in vehicle search. Existing datasets predominantly serve vehicle Re-ID from pre-cropped patches or single-task settings, whereas CLIPVehicle targets end-to-end vehicle search systems that must both detect vehicles and identify them across cameras in full scenes. In that benchmark design, CityFlowVS captures real surveillance characteristics and domain noise, SynVS-Day provides a day-only synthetic setting with reduced environmental confounds, and SynVS-All extends the synthetic setting to day, rain, dawn, and night (Wang et al., 6 Aug 2025).

Within that triad, SynVS-Day occupies an intermediate role. It has more data than the real-world set and fewer environmental confounds than SynVS-All. The dataset is therefore intended to isolate the core detection-plus-Re-ID difficulty while minimizing interference from adverse lighting and weather. The stated rationale is consistency with “most vehicle Re-ID datasets and CityFlowVS,” whose data are day-biased (Wang et al., 6 Aug 2025).

This positioning is methodologically important because vehicle search integrates two objectives that are not naturally aligned: detection emphasizes shared vehicle commonness, whereas re-identification emphasizes individual vehicle uniqueness. SynVS-Day functions as a controlled benchmark for studying that conflict without simultaneously introducing rain, dawn, or night variation (Wang et al., 6 Aug 2025).

2. Source corpus and construction procedure

SynVS-Day is constructed from Synthehicle, described as a massive synthetic multi-target multi-camera vehicle tracking dataset recorded in virtual cities. The source corpus contains 17 hours of videos from 340 cameras positioned around crossroads and highways across 64 scenes, and spans day, rain, dawn, and night conditions. The CLIPVehicle paper does not specify the rendering engine, shaders, or precise rendering settings (Wang et al., 6 Aug 2025).

The SynVS-Day subset is derived by processing the Synthehicle training towns in a way aligned with the CityFlowVS construction procedure. The reported steps are the following (Wang et al., 6 Aug 2025):

• Dropped pedestrian annotations: pedestrian boxes were removed.
• Dropped vehicles visible only in one scenario: this enforces cross-camera identity matching.
• Built splits by town: Town01, Town02, and Town03 are used for training, while Town04 and Town05 are used for testing.
• Isolated day scenes: only daytime scenes are retained to form SynVS-Day.

The resulting dataset inherits Synthehicle’s multi-camera layout and ground-truth annotations. Cross-camera variation comes from fixed, non-overlapping viewpoints around intersections and highways, producing viewpoint changes, scale variation, and occlusions typical of traffic scenes. Because vehicles appearing in only one scenario are excluded, identities in SynVS-Day are deliberately constrained to be observable across different cameras (Wang et al., 6 Aug 2025).

3. Dataset composition and annotation scope

The CLIPVehicle paper reports the following statistics for SynVS-Day (Wang et al., 6 Aug 2025).

Split	Identities	Frames / boxes / queries
Total	810	17,912 frames; 145,398 boxes; 1,277 query images
Train	475	11,102 frames; 95,473 boxes
Test	335	6,810 frames; 49,925 boxes; 1,277 queries

SynVS-Day provides bounding boxes for vehicles and cross-camera global identity labels represented as integers. Vehicles that appear in only one scenario are excluded, and pedestrian boxes are not included. Attribute labels such as color, type, make, and model are not annotated. Although CLIPVehicle uses attribute-aware textual prompts for identity discrimination, those prompts are learned without explicit attribute labels in SynVS-Day (Wang et al., 6 Aug 2025).

Because the dataset is synthetic, the annotations are inherited from the original tracking corpus and are exact rather than human-labeled. The paper describes them as per-frame box coordinates and per-instance global IDs. At the same time, several implementation-level properties are left unspecified: the exact camera count used in SynVS-Day is not reported, nor are image resolution, original frame rate, per-identity image distribution, file formats, or annotation schemas such as JSON or XML (Wang et al., 6 Aug 2025).

During training, raw images are resized to $900 \times 1500$ pixels. No shadow or illumination annotations are provided, although day lighting naturally introduces shadows in the virtual scenes (Wang et al., 6 Aug 2025).

4. Splits, search protocol, and evaluation metrics

The train/test split is town-wise, with Town01–03 used for training and Town04–05 used for testing. This enforces scene-level separation and non-overlapping environments. Identities do not overlap between train and test, following standard vehicle Re-ID practice (Wang et al., 6 Aug 2025).

The test protocol uses cropped query vehicle images against gallery frames composed of raw scenes. System outputs are detected boxes ranked by identity similarity to the query. For SynVS-Day, the paper reports the number of queries but does not detail the exact rule used to select them. It does specify that, in evaluation, a detected gallery box is counted as a true positive when its IoU with the ground-truth box of the same identity exceeds $0.5$ (Wang et al., 6 Aug 2025).

Vehicle search is evaluated with mAP and CMC Top-1. The benchmark does not use IDF1, Recall, or Precision. The reported metric definitions are (Wang et al., 6 Aug 2025):

$$\mathrm{mAP} = \frac{1}{N}\sum_{i=1}^{N} \mathrm{AP}_i$$

where $\mathrm{AP}_i$ is the area under the precision–recall curve for query $i$, and in discrete form,

$$\mathrm{AP}_i = \sum_{k=1}^{K} P_i(k)\,\Delta R_i(k).$$

For CMC, the benchmark reports Top-1:

$$\mathrm{CMC}@k = \frac{1}{N}\sum_{i=1}^{N} \mathbb{1}[\exists\, r \le k \text{ s.t. } \text{hit}_i(r)=1].$$

These definitions formalize an integrated detection-and-identification setting rather than a patch-only re-identification setting. Negative samples and distractors arise naturally from detections of other vehicles and background across all gallery frames (Wang et al., 6 Aug 2025).

5. Use in CLIPVehicle training

SynVS-Day is not only an evaluation set but also part of the benchmark used to train and assess CLIPVehicle as a unified vehicle search framework. CLIPVehicle combines a dual-granularity semantic-region alignment module with a multi-level vehicle identification learning strategy. The former uses vision-language modeling for vehicle discrimination, while the latter learns identity representations from global, instance, and feature levels (Wang et al., 6 Aug 2025).

The training objective jointly optimizes detection, re-identification, semantic-region alignment, and multi-level identification learning:

$\mathcal{L} = \mathcal{L}_{\mathrm{det} + \mathcal{L}_{\mathrm{reid} + \mathcal{L}_{\mathrm{sra} + \mathcal{L}_{\mathrm{mil}.$

Within that framework, the reported ingredients are (Wang et al., 6 Aug 2025):

• Detection loss $\mathcal{L}_{\mathrm{det}}$: a standard Faster R-CNN-style sum of foreground/background classification and smooth $L_1$ bounding-box regression, applied to preliminary and refined boxes.
• ID discrimination loss $0.5$0: Online Instance Matching operating on features from the Norm-Aware Embedding head.
• Dual-granularity semantic-region alignment $0.5$1: an object-granularity foreground/background alignment term and an ID-granularity alignment term using CoOp-learned tokens and attribute-augmented prompts.
• Multi-level identification learning $0.5$2: image-level multi-label BCE over global backbone features, box-level cross-entropy over ground-truth crops, and feature-level $0.5$3 distillation toward a frozen Re-ID model on ground-truth crops.

The dataset’s structure is directly compatible with these objectives. Every frame contains multiple vehicle identities, enabling multi-label image-level supervision, and the exact synthetic annotations support box-level and feature-level learning without manual labeling noise. The paper explicitly recommends training end-to-end search models with a shared detection-and-identification backbone and notes that methods designed solely for cropped-patch Re-ID degrade when they are driven by imperfect detections (Wang et al., 6 Aug 2025).

6. Reported performance and comparative position

On SynVS-Day, CLIPVehicle is reported to outperform both vehicle Re-ID methods adapted from patches and person-search-style joint detection-plus-identification methods. The representative results are as follows (Wang et al., 6 Aug 2025).

Method	mAP / Top-1
CLIPVehicle	32.4 / 84.1
CLIP-ReID	29.3 / 56.4
MSINet	18.6 / 35.9
MBR	16.3 / 80.1
SeqNet	30.2 / 82.6
COAT	31.6 / 83.0
OIMNet++	30.0 / 83.1

The paper interprets the gap over the strongest person-search baselines primarily through mAP, indicating better ranking quality across gallery detections. SynVS-Day also exposes broader dataset effects. CLIPVehicle’s mAP is higher on SynVS-Day than on CityFlowVS, where the reported mAP is $0.5$4, and higher than on SynVS-All, where the reported mAP is $0.5$5. The benchmark therefore captures both a synthetic-versus-real domain gap and an adverse-conditions gap (Wang et al., 6 Aug 2025).

The comparisons further clarify the intended use of SynVS-Day. Relative to CityFlowVS, it is synthetic, larger, and supported by exact ground-truth. Relative to SynVS-All, it is smaller and less environmentally variable, making it suitable for controlled experiments on joint detection and identification before robustness testing under rain, dawn, and night conditions (Wang et al., 6 Aug 2025).

7. Limitations, release status, and research significance

Several properties of SynVS-Day remain unspecified in the current paper. The graphics engine, rendering settings, per-identity image distribution, camera counts per town, frame rates, query selection procedure, synthetic photorealism, occlusion statistics, file formats, and annotation schemas are not detailed. The paper also does not provide a URL, license, or official file structure, stating only that the benchmark and code “will be released to the public” (Wang et al., 6 Aug 2025).

These omissions matter for reproducibility and downstream integration. A plausible implication is that some aspects of data loading, query formation, and synthetic-to-real transfer analysis will need to be verified against the eventual public release or against the Synthehicle source paper. Even so, the intended research use is explicit: SynVS-Day provides a controlled, day-only synthetic benchmark with substantial scale and cross-camera coverage for training and evaluating end-to-end vehicle search systems (Wang et al., 6 Aug 2025).

In that sense, SynVS-Day serves two complementary purposes. First, it is a development substrate for unified search models that must optimize detection and identity discrimination simultaneously. Second, it is an analytical control condition within the broader CLIPVehicle benchmark, allowing results on real surveillance video and mixed synthetic conditions to be interpreted against a cleaner day-only reference point (Wang et al., 6 Aug 2025).