The hidden danger of manhole cover detection based on KDWC-YOLOv5(Knowledge Distillation Well Cover-YOLOv5)

🧨 Congratulations! We achieve 0.948 of mAP on 350 test images in a competition!

To-do list:

• Model Ensembling ultralytics/yolov5#318
• Shape IOU https://github.com/malagoutou/Shape-IoU
• Multi-scale training
• Knowledge distillation https://blog.roboflow.com/what-is-knowledge-distillation/
• kepp training model, delving into best epoch and hyperparameter settings
• Publish dataets after the competition ends.
• Publish a web app made by my team.

Pre-requisties

• Linux

• Python>=3.7

• NVIDIA GPU (memory>=30G) + CUDA cuDNN

Strat evaluating

Install dependencies

pip install -r requirements.txt

Download the checkpoint and dataset

Our model‘s best checkpoint and dataset are located at the links below, you can download them freely.

Checkpoint: https://drive.google.com/file/d/1fclRgDYc_duWns63MbTeKRffmSPdP7BA/view?usp=sharing

Dataset: https://drive.google.com/file/d/16f29aRAM8zAsiaks8zuPdAzIkuEz1RKA/view?usp=sharing

Evaluation

If you want to get the mAP value, run the following command:

python val.py

If you want to get the images with bounding boxes, run the following command:

python detect.py

If you have a GPU cluster, I also provide you with a script file using sbatch to submit, and you can run it with:

sbatch yolov5_val.sh/sbatch yolov5_detect.sh

Tips: You can also use "--" to add parameters in the running command according to yourslef. E.g. If I want to output a txt file with the order of "Image name Confidence coefficient Coordinates", you can run the command below:

python detect.py --save-txt --save-conf

Training by yourself

If you want to train our model by yourself, you should firstly change the specify the path of your dataset in "data/A30.yaml", and you also nedd to specify a pretrained model, we use yolov5m, or you can choose other pretrained model via official link, then can run the following command:

python train.py

Tip1: You can also use "--" to add multi-scale parameters in the running command if you want to multi-scale training:

python train.py --multi-scale

Tip2: It is better to put the pretrained model under the root directory.

Web App Demo

Web_APP_demo.mp4

Wechat Mini Program Demo

1656_1715756946.mp4

⭐If you want to get the this app's developing codes or have any other questions, please feel free to conatact [email protected].