Small Object Detection

Small object detection in real-world construction site

This is a project for industial purpose and is in use now.

In real-world construction site, a lot of small objects, like safety halmets, need to be detected for safety construction. While these objects are too small for general detection algorithms. We develop a method to handle these hard samples.

We develop our algorithm based on FPN.

The baseline accuracy is about 60%, and I achieve an improvement of over 15%.