Road-scene object detection on BDD100K
MAUK R&D · computer vision
A road-scene detection model trained from scratch on 70,000 images with a laptop GPU, and an honest account of where it struggles.
A detection model for cars, people, signs and lights, trained on the public BDD100K driving dataset on consumer hardware, then tested on 10,000 unseen images and broken down by weather, scene and time of day.
- What it is
- Research
- Industry
- Cross-industry
- Status
- Research
- Year
- 2026
The problem
What needed solving
Overall accuracy hides the conditions that matter in the real world. The question wasn't only how good the model is, but where it fails: at night, in rain, on highways. And it had to be answered on hardware a small team actually owns.
The solution
What we built
A complete, repeatable path from raw dataset to trained model to a written report on where the model is strong and where it is weak.
Core functionality
- Dataset checks before any training starts
- Training that survives interruptions and picks up where it stopped
- Accuracy per object type, and where the model confuses one for another
- Accuracy broken down by weather, scene and time of day
- Curated prediction examples for each hard scenario
Screens
More screenshots
Captured from the live site, a demo workspace, or the product running on sample data. No customer records, passwords or private admin data are shown.
Business impact
What it achieved
0.575
detection accuracy (mAP50) on 10,000 unseen images
Verified0.330
strict detection accuracy (mAP50-95)
Verified−14.3 pts
hardest scene: highways vs residential streets
Highway 0.502 vs residential 0.645 mAP50.
Verified75.5 h
of training on a 6 GB laptop GPU
Verified
What sets it apart
Why it works
- 01
Where it fails, not just how well
Results are split by the conditions that matter on real roads, such as night, rain and highways.
- 02
Serious results on modest hardware
Trained from scratch on a 6 GB laptop GPU, which shows what a lean team can deliver without a data centre.
- 03
Nothing assumed about the data
The dataset was inspected before training, and categories that didn't belong were found and excluded.
Services
Where this work fits
Related projects
Similar projects
Prior work · client confidential
Computer vision
The client's system is not shown
Prior workSports ball-tracking computer vision
Prior work · national cricket body (confidential)
Computer vision that follows a small, fast ball across broadcast frames.
Machine learning
View project
Demo buildVerifiedCLAW & CO. AI virtual try-on
MAUK showcase store for beauty & nail brands
Shoppers see the set on their own hand before they buy.
AI agents · E-commerce · Machine learning
View project
LiveIn-house buildVerifiedMulti-disease risk screening model
MAUK R&D · applied machine learning
Risk scores for ten conditions from 25 routine measurements, trained on 281,000 records and private enough to run on the patient's own device.
86.88%accuracy on patient records the model never saw
Machine learning
View project
Discuss a similar project
Building something like Road-scene detection?
We'll tell you what we'd reuse, what we'd change and what it would take.