YOLOv3-Counter-Strike-Global-Offensive

This program is trained to detect and classify terrorists and counter-terrorists from a CS:GO gameplay clip. I made a dataset containing 434 images from CS:GO gameplay. Then I labelled the bounding box manually for each image and trained YOLOv3 model on the dataset.

Dependencies

• Tensorflow
• Keras
• Pillow
• Numpy
• Matplotlib
• colorsys
• OpenCV

Quick Start

1. Run convert.py to convert darknet weights to keras weights. Download darknet model from YOLO website.

2. Use create_annotationsV2.py to manually create annotations for images and save them to train.txt file. I have already done that boring part for you.

   • Press 1 to set class to '0' (CT).
   • Press 2 to set class to '1' (Terrorists).
   • '0' is the default class for every image.

   Example1: Example2:

3. Modify train.py and start training. I have already included a trained tiny_yolo model.

4. yolo_video.py uses trained model to detect objects in video/images.

To know more about issues and the YOLOv3 implementation that I used, refer to qqwweee/keras-yolo3. To learn the implementation of annotations tool, watch this tutorial mark-jay-yolo.

---

Result obtained using tiny_yolo model

As seen in above sample, model struggles to classify terrorists properly. Since it is obtained using tiny_yolo model, something like this was expected. I don't have a gpu to train full model.