Keras-retinanet-Training-on-custom-datasets-for-Object-Detection-

Prerequisites

1. Clone this repository.
2. Ensure numpy is installed using pip install numpy --user
3. In the repository, execute pip install . --user. Note that due to inconsistencies with how tensorflow should be installed, this package does not define a dependency on tensorflow as it will try to install that (which at least on Arch Linux results in an incorrect installation). Please make sure tensorflow is installed as per your systems requirements.
4. Alternatively, you can run the code directly from the cloned repository, however you need to run python setup.py build_ext --inplace to compile Cython code first.
5. Optionally, install pycocotools if you want to train / test on the MS COCO dataset by running pip install --user git+https://github.com/cocodataset/cocoapi.git#subdirectory=PythonAPI.

Note : Make sure that tensorflow is installed on the machine (tensorflow version < 2.0) ,python <=3.7

Prepare Data for training

1. Clone or download this repository.

2. Delete files inside snapshot,tensorboard,dataset/annotation and dataset/images in this directory.

3. Paste all training images in dataset/images directory.

4. Next step is to create annotation file for each images and place all annotation xml files in dataset/annotations directory. Which will be in xml format. LabelImg is one of the tool which can be used for annotation. LabelImg github or LabelImg exe

5. Then have to set the config file custom_dataset_config.py inside config directory. Here set the path for annotation, image, train.csv files and also set the path where the classes.csv have to be saved. TRAIN_TEST_SPLIT value will split the data for training and testing.

6. create train.csv, test.csv and classes.csv by using build_dataset.py inside this directory. Run the following command inside the directory

    `python build_dataset.py`
   

   The datas inside train.csv and test.csv will be in the following format.

    `<path/to/image>,<xmin>,<ymin>,<xmax>,<ymax>,<label>`
   

   classes.csv will be in following format.

     `label, index`
   

The data is all set for training. Now lets start Training.

Training

Download the pretrained weights on the COCO datasets with resnet50 backbone from this link. Paste it in the directory.

Start training by run the following one of the command.

$ retinanet-train --weights resnet50_coco_best_v2.1.0.h5 --steps 400 --epochs 20 --snapshot-path snapshots --tensorboard-dir tensorboard csv dataset/train.csv dataset/classes.csv --val-annotations dataset/test.csv

or python keras_retinanet/bin/train.py --weights resnet50_coco_best_v2.1.0.h5 --steps 400 --epochs 20 --snapshot-path snapshots --tensorboard-dir tensorboard csv dataset/train.csv dataset/classes.csv --val-annotations dataset/test.csv

After each epoch, the model will be saved in the snapshots folder. Stop the training by using 'ctr+c' when the loss will be minimum.

Convert Model

The model have to be converted in a format which can be used for prediction. For this run one of the following command.

$ retinanet-convert-model <path/to/desired/snapshot.h5> <path/to/output/model.h5> or

$ python keras_retinanet/bin/convert_model.py <path/to/desired/snapshot.h5> <path/to/output/model.h5>

Note : If you set any anchor params (--config) during training, then you should pass it to the above command while model convertion. (eg : retinanet-convert-model <path/to/desired/snapshot.h5> <path/to/output/model.h5> --config config.ini). Refer the 53 line (parser.add_argument('--config', help='Path to a configuration parameters .ini file.')) in keras-retinanet/bin/convert_model.py

Start Detection

Set model_path, video_path, output_path and labels_to_names values in the people_detection_video.py Run the following command.

python people_detection_video.py

An output video will be saved in the mentioned path.

Also detecting from image is done by run the following command.

python people_detection_image.py

Output of my object detection model.