A PyTorch implementation of RetinaNet, with support for training, inference and evaluation.

The method of RetinaNet was used to perform defect detection on NEU surface defect database, and We adopted data enhancement methods such as random clipping, flipping and color enhancement. Finally achieved a satisfactory result.

$ git clone https://github.com/Gmy12138/RetinaNet.git
$ cd RetinaNet/
$ pip install -r requirements.txt

$ cd scripts/
$ run get_state_dict.py

$ Download address    http://faculty.neu.edu.cn/yunhyan/NEU_surface_defect_database.html
$ cd data/
$ Put the dataset in the data folder

Evaluates the model on NEU-DET test.

Uses pretrained weights to make predictions on images. The ResNet-50 measurement marked shows the inference time of this implementation on GPU 2080ti.