YOK: Yolov2 On Keras

Features

• Object detection in a image/video using yolov2 based on keras framework(tensorflow backend).
• It includes training and inference.
• Yolov2 and TinyYolov2 models can be trained.
• It is tested using COCO dataset.

Requirements

• Keras
• tensorflow
• scipy
• matplotlib
• numpy
• tqdm
• opencv-python
• baker
• cytoolz
• lxml
• pandas
• pillow
• imgaug

Note: Please take note for tensorflow gpu support, install the cuda driver, cuda toolkit and cudnn from this [link] (https://docs.nvidia.com/deeplearning/sdk/cudnn-install/index.html) and the corresponding tensorflow-gpu version supported by refering to this link

Inference/Prediction

• Please download the pre-trained models from this link
• Alternatively one can re-train the model files as described in the Training section
• Move the weights file under h5_models/pre-trained
• Adjust model to use Yolov2/TinyYolov2, pre-trained weights file path and the class to target/detect in config_predict.json
• Run the script PYTHONPATH=lib:$PYTHONPATH CUDA_VISIBLE_DEVICES="" python3 predict.py (You can adjust the GPU devices to be used using CUDA_VISIBLE_DEVICES)
• By defualt the video under videos/conference.mp4 is used as input and the output video is saved as videos/conference_detect.avi and "class=person" is used for detection
• The csv file contatining the number of person per second of the video is saved as logs/number_count.csv
• One can input an image file too using command PYTHONPATH=lib:$PYTHONPATH CUDA_VISIBLE_DEVICES="" python3 predict.py -i images/person.jpg -o output_images/person.jpg
• Run PYTHONPATH=lib:$PYTHONPATH python3 predict.py --help to see script arguments and defaults.

Output

Training

• For training, pre-trained weights files (h5 models) are used.
• Please download the pre-trained models from this link
• Move the weights file under h5_models/pre-trained
• Download the COCO dataset You can create annotations using the two following ways:
  • Convert COCO annotations to VOC format:
    • Train set : http://images.cocodataset.org/zips/train2014.zip
    • Validation set: http://images.cocodataset.org/zips/val2014.zip
    • Annotations : http://images.cocodataset.org/annotations/annotations_trainval2014.zip
    • Extract train2014.zip and val2014.zip and move the train2014 and val2014 directory under data/coco/images
    • Extract annotations_trainval2014.zip and move the annotations_trainval2014 directory under data/coco
    • Go to data directory and run the script python3 coco2pascal.py
    • The annotations will be saved under data/coco/pascal_format (It will take considerable amount of time)
  • Download the VOC annotations for COCO dataset:
    • Download the annotations from this link
    • Unzip pascal_format.zip and move the directory pascal_format under data/coco
• Adjust epochs, batch_size in config_train.json, also the pre-trained weights file path and which model to use Yolov2 or TinyYolov2
• Also specify the train and eval data path in config_train.json
• Run the script PYTHONPATH=lib:$PYTHONPATH CUDA_VISIBLE_DEVICES="" python3 train.py (You can adjust the GPU devices to be used using CUDA_VISIBLE_DEVICES)
• The output model file is saved in h5_models/trained
• Run PYTHONPATH=lib:$PYTHONPATH python3 train.py --help to see script arguments and defaults.