Danil Sidorov
RU EN DE
Telegram

Dataset for an autopilot

A labeled road sign dataset for training an autopilot that won a competition

AI / Компьютерное зрение

Live interactive demo stand

An exact copy of the project interface and business logic with demo data. All actions run in real time.

43 800 размеченных вручную кадров, 89 классов ГОСТ, победа в соревновании

StackPython, Computer Vision, YOLO, PyTorch, Hugging Face

Safe perimeter (NDA): production logic on test data

What to test in the stand:
  1. Переключите дорожные сцены с автомобильного видеорегистратора
  2. Включите тумблер YOLO-разметки с подсветкой классов и уверенности модели
  3. Изучите нормализованные координаты и класс объекта в инспекторе датасета
https://cv-demo.internal/dataset/inspect
Live Stand
Road Signs Vision YOLOv8 Detection Inspector · 89 GOST Classes
TensorRT FP16 · 14.2 ms / frame Hugging Face Hub · 43.8k frames
CAM_01 · FRONT 120° FOV FRAME_ID: #02941
Детектированные объекты в кадре
2 объекта
YOLO-разметка кадра (labels/*.txt)
Формат: <class_id> <x_center> <y_center> <w> <h>

                

Turnkey implementation

Need a similar system for your business processes?

I will develop the architecture, fault-tolerant backend, and fast interface. Protected by NDA, clean transferable code, and transparent milestones.

  1. Free timeline and budget estimate within one day
  2. Contract and invoice
  3. Warranty and post-launch support

On weekdays usually within a couple of hours

I will reply personally, no aggressive sales pitch

Task

There was not enough data with the right set of classes to train a road sign detector. Open datasets did not fit in labeling or in the set of signs.

Solution

I collected and labeled a road sign dataset, selected the images, brought them to one format and annotated them by class. The structure fits a standard training pipeline, so the data works without relabeling.

What was done

  • Collected and selected road sign images for the required classes
  • Brought the images to one format
  • Annotated the images by class
  • Organized the structure for a standard training pipeline
  • Trained a semi-automatic labeling tool
  • Doubled the dataset with the labeling tool
  • Post-processed the data and handed it over to the client

Result

A ready dataset that a sign detector trains on right away. The set of classes grows as needed. An autopilot trained on this dataset won competitions. Links to them are below.

Technologies

Computer VisionDataset

Sources

The links confirm the numbers in this case

Tell me about your task

Telegram