yolov5 for semantic segmentation which based on flexible-yolov5.

• Features
• Prerequisites
• Getting Started
  • Dataset Preparation
  • Training
  • Visualization
• Reference

• Provide model structure, such as backbone, neck, head, can modify the network flexibly and conveniently
• backbone: yolov5s
• neck: FPN, PAN
• head: segmentation head

please refer to requirements.txt

1. Download coco or cityscapes dataset.
2. Modify your dataset path in configs/data.yaml.

There are some modifies in classes grouping with cityscapes dataset.

• Change: 19 classes ——> 9 classes (See the code details in cityscapes.py. But don't worry, I have add the input param group(bool) which can be set as False to backend 19 classes in training and visualizing. )

For training, it's same like yolov5.

You can modify your setup in train_cityscapes.sh.

$ ./train_cityscapes.sh

Tensorboard is automatically started while training. You can see the visualizition results in tensorboard.

You can modify your setup in visualize.sh.

$ ./visualize.sh

Here is the visualization of cityscapes with 9 classes(ground truth is 19 classes).

• ultralytics/yolov5