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.
- Tensorflow
- Keras
- Pillow
- Numpy
- Matplotlib
- colorsys
- OpenCV
-
Run
convert.py
to convert darknet weights to keras weights. Download darknet model from YOLO website. -
Use
create_annotationsV2.py
to manually create annotations for images and save them totrain.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.
-
Modify
train.py
and start training. I have already included a trainedtiny_yolo
model. -
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.
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.