A keras FRCNN to detect Fox in Super Smash Bros Melee for the Nintendo Gamecube. The FRCNN object detector was written by yhenon and the orignal repo can be seen here. The gameplay used to train and test the model is from VGBootcamp (twitch & youtube). The model was trained on the AMI provided by PyImageSearch.
This google drive folder has the trained model file, model_frcnn.hdf5. The folder also has the training data, data.zip, which contains a folder of images in addition to bounding box annotations produced by dlib's imglab (these annotations were converted to fox_frcnn_tags.txt for compatibility with the frcnn scripts). Lastly the folder has a .zip containing images with bounding boxes drawn by the trained model, results_imgs.zip.
This gif shows the output of the model. The model works with still images 1 at time; a sample of these single frame outputs were then converted into the gif.
If you'd like to use the trained fox model you'll need to:
- clone this repo
- download
model_frcnn.hdf5to the cloned directory from here - create a subdirectory named
results_imgsin the cloned directory - run the command:
python test_frcnn.py -p /path/to/folder/with/images/to/annotate
The model will output the annotated images to results_imgs.
Note: the model was trained/tested only on blue (purple) melee fox