SimVeRi: a synthetic vehicle re-identification dataset with spatiotemporal metadata and aerial–ground views

Overview SimVeRi is a synthetic multi-camera vehicle re-identification dataset generated through synchronised CARLA–SUMO co-simulation for roadside, elevated, and aerial observation settings. The release contains 70,773 unique vehicle-image contents across the core roadside-camera benchmark, the Twins supplement, and the aerial–roadside (AG) protocol resources. The dataset provides structured, component-specific metadata linking images to vehicle identity, camera identity, simulation timestamps, world coordinates, vehicle headings, and motion states. Bounding-box, occlusion, visibility, and other quality fields are provided where applicable and are documented in the accompanying metadata schema and README. SimVeRi is intended as a controlled, metadata-rich testbed for reproducible vehicle ReID evaluation, spatiotemporal reasoning, controlled appearance ambiguity analysis, trajectory-aware association, and heterogeneous aerial–roadside vehicle matching. Dataset Components 1. Core Roadside-Camera Benchmark (legacy package label: GG) - 33,942 images of 650 vehicle identities- Captured by 24 fixed non-air cameras (19 roadside and 5 elevated cameras)- Identity-disjoint train / gallery / query split (454 train IDs, 196 eval IDs; 28,379 train, 4,738 gallery, 825 query images)- One query image per camera per vehicle; query excluded from gallery 2. Twins Supplement - 14,032 images of 125 vehicles in 25 visually identical groups (5 vehicles per group)- All vehicles within each group share the same CARLA blueprint and colour- Supports controlled evaluation under severe appearance ambiguity at both vehicle-level and group-level 3. Aerial–Roadside Extension (legacy package label: AG) - Tracklet-level cross-view protocols for aerial–roadside vehicle association- Test scope: 316 aerial tracklets / 825 roadside tracklets (79 vehicles with both layers)- Full scope: 984 aerial tracklets / 1,077 roadside tracklets (246 vehicles with both layers)- Positive pairs defined by same vehicle identity and temporal overlap- Aerial and roadside tracklet images additionally include per-image spatiotemporal metadata (ag_spatiotemporal.json: timestamp, world coordinates, heading, speed, occlusion).- The complete release contains 70,773 unique image contents.- The legacy AG protocol directories contain replicated copies of selected core and aerial images; protocol-level file counts are therefore not additive.- Resource-qualified relative paths, rather than basenames alone, should be used as global image identifiers. Component-specific image metadata Metadata availability differs slightly between released components. Depending on the resource, image records include: vehicle identity and camera identity; simulation timestamp and frame identifier; world coordinates in the CARLA coordinate system; vehicle heading and speed; bounding-box dimensions and camera-to-vehicle distance; occlusion, visibility, and other quality attributes where available; Twins group identity for controlled appearance ambiguity analysis. The exact fields and their resource-specific availability are documented in the README and metadata schema. Licence-Plate Rendering and Identity Cues The native CARLA vehicle assets render licence plates as fixed surface textures. The generation pipeline does not assign vehicle-specific registration strings or randomise, blur, or mask plate textures. A structured audit examined 200 front- and rear-view images across all 25 released vehicle blueprints. All legible sampled plates displayed the placeholder text “CARLA”, and no identity-specific plate text was observed. Plate-region pixels remain part of each vehicle image and may contribute blueprint-level appearance. No plate-masking ablation was conducted, so the influence of the plate image region was not separately estimated. Within each Twins group, plate appearance is identical because all members share the same blueprint and colour. Dataset-Level Metadata - camera_network.json --- 30-camera deployment with 3D positions, rotations, FOV, layer classification, and route-based inter-camera distance matrix- trajectory_info.csv --- within-camera tracklet summaries (2,731 tracklets)- camera_transitions.csv --- inter-camera transition statistics (route distance, mean travel time, sample counts)- splits.json --- identity-disjoint train/test partition (seed 42, ratio 0.70)- vehicle_attributes.json --- vehicle type and colour definitions- twins_groups.json --- group membership for identical-appearance vehicles- statistics/image_manifest.csv — release-wide image manifest containing resource namespace, relative path, SHA-256 hash, content ID, replicated-copy indicator, and filename-collision indicator- SHA256SUMS — SHA-256 checksums for the root-level release files and archives Intended Use SimVeRi supports research in: - Vehicle re-identification and multi-camera retrieval- Spatiotemporal reasoning and trajectory-aware association- Cross-view and aerial–roadside vehicle matching- Controlled evaluation under appearance ambiguity (Twins)- Synthetic data for intelligent transportation systems- Camera network analysis and deployment optimisation Generation SimVeRi was generated using CARLA 0.9.13 and SUMO 1.15.0 in synchronised co-simulation mode. Vehicle appearance was sampled from 25 native CARLA blueprints with 17 colour variants. The simulator weather was fixed to ClearNoon. The companion generation and validation code is available at: https://github.com/Zengzhi-Zhang/SimVeRi