Team: Broncos

Team members: Tianjie Zhang, Donglei Wang, Yang Lu

Email: [email protected]

🚗Task 1

How to run:

0. Download this repository and put it in your tutorial file folder.

1. open the terminal and go to your file folder:

cd 'your_path'

2. Run the following code to train the mdoel.

python train_dsps.py --data data/dsps.yaml --epochs 400 --weights yolov5s.pt --cfg yolov5s.yaml  --batch-size 16 --hyp data/hyps/hyp.scratch-med.yaml

3. After training, run the detec.py to test the result.

python detect_dsps.py --weights 'your_trained .pt file path' --source test_data --conf-thres 0.65 --iou-thres 0.999 --augment

Other way: run the DSPS_task1.ipynb directly.

Strategy used:

1. Using wGAN to generate real-like images;
2. using random crop, flip and contrast adjusting to augment the images;
3. A feature-balanced strategy: try to make the amount of different defects equal'
4. Fine tune of the hyperparameters: training: epochs = 400, batch_size = 16, hyp = hyp.scatch-med.yaml testing: conf-thres = 0.65, iou-thres = 0.999, augment = true

---

🚓Task 2

How to run:

1. open the terminal and go to your file folder:

cd 'your_path'

2. Run the following code to train the mdoel.

python train_dsps.py --data data/dsps2.yaml --epochs 400 --weights yolov5s.pt --cfg yolov5l_simAM.yaml  --batch-size 16 --hyp data/hyps/hyp.scratch-med.yaml

3. After training, run the detec.py to test the result.

python detect_dsps2.py --weights runs/train/exp10/weights/best.pt --source images/ --conf-thres 0.65 --iou-thres 0.999 --augment  

Other way: run the DSPS_task2.ipynb directly.

Strategy used:

A sinAM Attention module was used to modify the yolov5l model.

---

Links:

• https://dsps-1e998.web.app/
• https://github.com/UM-Titan/DSPS23