vehicle-research/multi-weather_traffic_data

VN2V-Weather

A multi-weather traffic vehicle re-identification dataset collected in Ho Chi Minh City, Vietnam, in May 2026, from two paired camera views of the same traffic scene. Both cameras were positioned on a pedestrian footbridge over a national highway: one faces approaching traffic (before, front view) and one faces departing traffic (after, rear view). The observed carriageway carries traffic in a single direction, so every vehicle crosses the scene in the same order.

The two streams cover the same traffic window but are not hardware-synchronized. Cross-view correspondence is established through manual identity annotation, not timestamp alignment.

Each annotated vehicle has a bounding box, a vehicle class, and an identity ID. The same physical vehicle receives the same identity ID in both views.

Conditions

Condition	Views	Description
morning_norain	before, after	Morning traffic without rain
morning_rain	before, after	Morning traffic with rain
evening_norain	before, after	Evening/night traffic without rain
evening_rain	before, after	Evening/night traffic with rain

Statistics

42,254 frames and 100,952 annotated boxes across eight streams.

Condition	View	Frames	Boxes	IDs
morning_norain	before	4,923	12,784	628
morning_norain	after	5,207	10,785	571
evening_norain	before	4,597	10,798	567
evening_norain	after	5,037	10,631	537
morning_rain	before	6,671	18,445	669
morning_rain	after	6,700	16,074	618
evening_rain	before	4,561	12,807	635
evening_rain	after	4,558	8,628	581
Total		42,254	100,952

Cross-view identity consistency:

Condition	Shared IDs	Before-only	After-only	Label mismatches
morning_norain	571	57	0	0
evening_norain	537	30	0	0
morning_rain	618	51	0	0
evening_rain	581	54	0	0

Class distribution: truck 48,801 (48.34%), motorbike 30,360 (30.07%), car 19,439 (19.26%), bus 2,352 (2.33%).

Identities are scoped to a condition. A vehicle in morning_rain and a vehicle in evening_rain never share an identity, so matching is defined between the two views of one condition and not across conditions. The 2,307 cross-view-matchable identities are the sum of the four Shared IDs values.

Format

Frames are 1080 x 1920 (29,842 frames) and 1440 x 2560 (12,412 frames). The no-rain sessions are entirely 1080p, the evening-rain session entirely 1440p, and the morning-rain session contains both. Each XML <image> element records its own resolution, so box coordinates always refer to the pixel grid of their own frame — no global rescaling is needed.