This repo uses official implementations (with modifications) of YOLOv7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors and Simple Online and Realtime Tracking with a Deep Association Metric (Deep SORT) to detect objects from images, videos and then track objects in Videos (tracking in images does not make sense)
I have refactored the code, removed dependencies, removed extra code so that you can use ANY detector model with DeepSORT. Please look at the Demo.ipynb notebook on how to use the code.
To use in Colab: Open Colab Demo.ipynb
For use in local system, please follow the below steps
- Clone the repo as
git clone https://github.com/deshwalmahesh/yolov7-deepsort-tracking/ - Download the weights of any of the pre trained
YOLOv7models from the links:yolov7.ptyolov7x.ptyolov7-w6.ptyolov7-e6.ptyolov7-d6.ptyolov7-e6e.pt - NOTE: Every model has it's own parameters like
image_sizeand all so you have to use the appropriate parameters. This repo was tested successfully withyolov7x.pt - Go to
Demo.ipynband run the code.
This code works perfectly with python== 3.7, tensorflow==2.8.0, torch== 1.8.0, sklearn==0.24.2 on local Ubuntu: CPU as well as Colab: CPU + GPU as of 13/07/2022.
One of the most frequent problem is with the PATH such as model weights, input, output etc so pass in the path of the weights carefully. Do not just run all all the cells given in the notebook. this code works perfectly as long as you pass the correct path.
If you find yourself in trouble, please raise an issue.
- YOLOv7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors
- Simple Online and Realtime Tracking with a Deep Association Metric (Deep SORT)
- Tensorflow-v1 error code help
Any issues, help, bug, feature requestd andd suggestions are very much welcomed. Please feel free to open up the issues.
I'll be putting the quantized Onnx deployable version here along with Docker image in some time. If you have the time and expertise, please free to open up a merge request.