Improving YOLOv3

Accompanying code for Paperspace blog Improving YOLOv3.

Dataset Preparation

Use Microsoft VoTT tool to label objects.

Download pretrained weights

$ mkdir weights
$ cd weights/
$ curl https://pjreddie.com/media/files/yolov3.weights --output yolov3.weights

Convert labelled data to YOLO format

$ python Convert_To_YOLO_Format.py path_to_exported_files_from_the_image_tagging_step_with_VoTT

Create yolo config file for any number of classes

$ python custom_model.py --num_classes 6 --file_name ./config/yolo-custom-6class.cfg

Training

Use pretrained weights to finetune the YOLOv3 model using tricks mentined in Improving YOLOv3 on your data.

$ python -m torch.distributed.launch --nproc_per_node=1 train.py --pretrained_weights "./weights/yolov3.weights" --n_cpu 1 --ngpu 1 --distributed True --model_def ./config/yolo-custom-6class.cfg

Credit

Bag of Freebies for Training Object Detection Neural Networks

Zhi Zhang, Tong He, Hang Zhang, Zhongyue Zhang, Junyuan Xie, Mu Li

Learning Spatial Fusion for Single-Shot Object Detection

Songtao Liu, Di Huang, Yunhong Wang

Most of the code in this repo has been adapted from here and here.