Dataset Methods Results Gallery Human Eval Download

ConSynth-X
Evaluating Construction AI Under Extreme Conditions

A large-scale dataset that introduces systematic, controllable environmental and operational variations to reflect the complexity of real construction workplaces, including diverse weather conditions (e.g., fog, rain, and snow), lighting variability, and scale differences.

34,199
Total Rows
16
Object Classes
11
Conditions
3
Sub-datasets
Explore Dataset GitHub Hugging Face
Excavator and workers in dense fog
Fog
Excavator and workers under heavy rain
Rain
Construction site under heavy snow
Snow
Construction site at night
Night
Open Gallery
Sections

Explore the Project

Release note · ConSynth-X official repository and dataset v1 went online on May 1, 2026. More information coming soon.

01 · Dataset

Data & Taxonomy

16 unique object classes across 34,199 rows in 3 sub-datasets with 11 synthetic conditions.

View Dataset → 02 · Methods

Generation Pipelines

Five synthetic pipelines for controlled condition generation.

View Methods → 03 · Gallery

Dataset Browser

Samples across diverse conditions (e.g., rain, snow, fog, night, scale/distance).

Open Gallery → 04 · Results

Validation & Findings

FID/KID distances, weather classifier accuracy, and texture belief fusion.

View Results →
00 — Motivation

The Robustness Gap

Construction sites operate under varying weather, lighting, and scale conditions that challenge vision system reliability, but most vision models are trained primarily on favorable-condition data.

ConSynth-X turns that mismatch into a measurable dataset with controlled environmental and operational variations.

Excavator and workers in dense fog at a construction site
Dense fog reduces contrast, masks distant workers, and makes equipment boundaries harder to localize.
The Robustness Gap
46% mAP drop under weather (0.819 → 0.446 on SODA)
45% mAP drop at nighttime (0.819 → 0.454 on SODA)
Note: Results based on YOLOv8 cross-condition evaluation on SODA.