Researchers release ConSynth-X, 34,199 synthetic jobsite images for testing site cameras in rain, fog and darkness

Four researchers at Kent State University and the University of North Texas posted a paper to arXiv on September 21, 2026, that releases ConSynth-X, a set of 34,199 construction-site images built to test jobsite cameras in bad weather and darkness. The team took 3,109 daytime photos from three public construction image collections and generated 11 altered versions of each, covering rain, snow, fog, night, precipitation at night, and distant views where workers and machines appear small.

The work applies to the computer vision behind camera-based safety alerts and progress tracking, where software reads site photos or video to find workers, hard hats and equipment. The paper checks whether the altered images look like real bad-weather photos, and it reports no accuracy results for any detection model run on them.

What it doesEvery one of the 34,199 images keeps the labels of its source photo

Most public jobsite images were shot in clear daylight, the authors say, because rain and night shifts are hard to photograph on a schedule and every image has to be labeled by hand. So a model that scores well on those images has not been tested on a worker without a hard hat before sunrise in fog. ConSynth-X gives a researcher the same scene under each condition. If accuracy drops on the fog version of a photo, the fog caused it, since the workers and the camera angle did not change.

The inputs are three existing datasets, and every altered image keeps the labels of the photo it came from. A bounding box is the rectangle drawn around each object a model is supposed to find, and ConstructionSite 10k supplies boxes for excavators, rebar and workers in white hard hats, along with safety-rule annotations. Because the labels carry over, nobody had to relabel 34,199 pictures.

How each condition was madeFive pipelines produced the 11 conditions

1. Rain and snow. An image-editing model rewrote the sky and surfaces from a text prompt, and a particle overlay then drew streaks or flakes in near, middle and far layers.
2. Fog. A depth model estimated how far each pixel sits from the camera, and a standard atmospheric formula faded distant objects more than near ones. Visibility was set at 750 to 1,000 meters for light fog, 500 to 750 for moderate and 300 to 500 for heavy.

3. Night. A day-to-night translation model relit each scene.
4. Rain or snow at night. The weather edit ran first, then the night translation, then the overlay.
5. Distant views. Each photo was shrunk inside a larger canvas, and an image model filled the border with ground and sky, so every object takes up less of the frame.

The authors compare their release with five earlier construction image sets. The largest, MOCS, has 41,668 real photos. ConSynth-X is the only one in their comparison with each synthetic image tied to a specific source photo.

Evidence so farAt a 0.90 similarity cutoff, 688 scenes would have survived instead of 3,109

The validation answers two questions. The first is whether each altered image still shows the original scene, and the second is whether it looks more like a real bad-weather photo than the original did. No detection model was scored on the new images, so this paper does not tell you how much accuracy a site camera loses in fog.

For the first question, the authors ran every image through DINOv3, an image model that turns a picture into a list of numbers describing what it shows. They compared each altered image with its source by cosine similarity, a score where 1.0 means the two images show the same content. A source scene stayed in the dataset only if all eight gated versions scored 0.85 or higher.

At 0.85, 3,109 scenes survived. At 0.90, only 688 would have. The heavy rain, rain at night and snow at night versions were kept without the gate, and their similarity scores ship in the metadata so users can filter them.

Realism81.85% of the altered images sat closer to real weather photos than their sources did

For the second question, the authors measured distance to real adverse-weather photos from three public sets, ACDC, WeatherBench and WeatherNet. They added 199 construction-site rain photos retrieved through the Openverse API (a search service for openly licensed images). On the ConstructionSite 10k subset, 81.85% of the altered images sat closer to the real-weather photos than their daytime sources did. Nine of the ten conditions they checked moved in that direction. Rain at night moved the other way.

The chart below uses FID, or Fréchet Inception Distance, a score of how different two sets of images look to a standard image-recognition network. Lower means more alike. Each row compares the original daytime photos and the converted photos against the same ACDC reference set.

Night conversion made the biggest difference. FID fell from 282.2 for the daytime originals to 191.4. Heavy rain went the wrong way, rising from 199.3 to 237.6, and the authors call heavy rain and precipitation at night inconsistent across reference sets. Apart from the 199 Openverse images, the reference photos come from road and weather collections outside construction.

CaveatsThe paper and the GitHub README disagree on two points

Two parts of the release disagree with each other, and both matter if you plan to use it.

The Hugging Face listing does show Parquet files for all 11 conditions, behind a gate that requires manual approval.

How to measure it on your projectsSplit camera recall into clear day, rain, fog and after dark

The paper's design, the same scene under different conditions, works as a test on your own footage. If you pilot an AI camera for PPE or exclusion-zone alerts, track detection recall by condition. Recall here is the share of real events a person finds on review that the system also flagged, split into clear day, rain, fog and after dark.

Capture the baseline before relying on the alerts:

1. Pick one fixed camera on an active job, and record the weather and light for each shift in the daily log.
2. Each week, have the safety manager review a set amount of footage from a clear shift and the same amount from a rain, fog or night shift, and log every PPE miss and zone entry.
3. Match the logged events against the system's alerts. Recall for a condition is flagged events divided by logged events.

If recall after dark runs well below the clear-day figure, those are the hours when a person still has to watch the feed. A first step this week is to add a weather-and-light field to the daily log for the area that camera covers.

Availability and costFree for noncommercial use under CC BY-NC 4.0

The dataset is free under the Creative Commons Attribution-NonCommercial 4.0 license, and the source datasets' own terms still apply. The license excludes commercial use, so a vendor would need separate permission to train a paid product on it. The generation code is on GitHub under the Apache 2.0 license, and users have to download the third-party model checkpoints from their own publishers. The work was funded in part by the Farris Family Innovation Award, with computing time from the Ohio Supercomputer Center.