Keras-RetinaNet Custom Object Detection

Installation process of keras-Retinanet

• Download the Keras-Retinanet source from https://github.com/fizyr/keras-retinanet
• Install Tensorflow > 2.3.0. In my case I have Tensorflow 2.3.1
• Execute pip3 install . –user.
• To compile Cython code run python setup.py build_ext –inplace

Test Keras-Retinanet using pre trained model

• Download pre-trained model of Keras-Retinanet from https://github.com/fizyr/keras-retinanet/releases and save it to snapshots folder.
• Go to examples folder and open ‘’’ resnet50_retinanet.py and confirm your downloaded model name inmodel_path = os.path.join('..', 'snapshots', 'resnet50_coco_best_v2.1.0.h5')```.
• After confirmation run python resnet50_retinanet.py. If it runs successfully, congratulation you have successfully installed Keras-Retinanet.
• If there is an compute_overlap error

File “..\keras_retinanet\utils\anchors.py”, line 20, in <module> 
from ..utils.compute_overlap import compute_overlap
ImportError: DLL load failed while importing compute_overlap

Than open “keras_retinanet\utils\ anchors.py” in your editor and edit

from ..utils.compute_overlap import compute_overlap

to

from compute_overlap import compute_overlap

This will solve the error of cumpute_overlap.

Train our custom dataset using keras-Retinanet

• Step1 : Prepare Dataset
• Step2 : Labelling/Annotation
• Step3 : Train
• Step4 : Test

Prepare Dataset

First, you need to define what object you want to detect. In this example i have used "racoon" images to train.[https://public.roboflow.com/object-detection/raccoon]. Copy all your images in “dataset/JPEGImages” folder.

Labelling/Annotation

The next step is to Label/annotate the dataset according to respective class. In this case there is only one class : "raccoon". I used labelImg from https://github.com/tzutalin/labelImg. It is a graphical image annotation tool written in Python and it uses Qt for its graphical interface. In labelImg confirm the save directory of “.xml” file to “dataset/Annotations” folder.

I have used PASCAL VOC format since it is supported by Keras-Retinanet.

Images are in “dataset/JPEGImages”

XML file are in “dataset/Annotations”

Divide the whole dataset into training and validation

Open/Run “train_val_divide.py” in and enter your desired percentage_val default is “percentage_val = 10;” . This will save “train.txt” and “val.txt” in directory “ImageSets/Main/”

Directory Structure of dataset

+---dataset
|   |   train_val_divide.py
|   |   
|   +---Annotations
|   |       raccoon-1.xml
|   |       raccoon-10.xml
|   |       raccoon-100.xml
|   |       …
|   +---ImageSets
|   |   \---Main
|   |           train.txt
|   |           val.txt
|   |           
|   \---JPEGImages
|           raccoon-1.jpg
|           raccoon-10.jpg
|           raccoon-100.jpg
|           …

Define the classes to "keras_retinanet/preprocessing/pascal_voc.py" Default is

voc_classes = {
    'aeroplane'   : 0,
    'bicycle'     : 1,
    'bird'        : 2,
    'boat'        : 3,
    'bottle'      : 4,
    'bus'         : 5,
    'car'         : 6,
    'cat'         : 7,
    'chair'       : 8,
    'cow'         : 9,
    'diningtable' : 10,
    'dog'         : 11,
    'horse'       : 12,
    'motorbike'   : 13,
    'person'      : 14,
    'pottedplant' : 15,
    'sheep'       : 16,
    'sofa'        : 17,
    'train'       : 18,
    'tvmonitor'   : 19
}

Changed to

voc_classes = {
    'raccoon'   : 0
    
}

Because i have only one class

Train

python keras_retinanet/bin/train.py --weights snapshots/resnet50_coco_best_v2.1.0.h5 --steps 1000 --epochs 10 --snapshot-path snapshots --tensorboard-dir tensorboard pascal dataset

--weights -> Initialize the model with weights from a file

--steps -> Number of steps per epoch

--Epochs -> Number of epochs to train

--snapshot-path -> Path to store snapshots of models during training

--tensorboard-dir -> Log directory for Tensorboard output

Pascal -> Directory of our dataset

Test

To test our trained model, first we need to convert it from training format to test format.

python keras_retinanet/bin/convert_model.py snapshots/resnet50_pascal_01.h5 directory_to_save_test_model.h5

Now we can this Saved model for our object detection. ** You can use "examples/resnet50_retinanet.py" to test the trained model

**Ref {"https://github.com/fizyr/keras-retinanet"]