Single Shot Object Detection in Tensorflow.Keras

This repository implements a single shot object detection framework in Keras.

Features

• Training
• Testing/Inference
• Custom Networks
• Custom Datasets
• MAP Evaluation
• Restore Training (Not fully implemented.)

Data Structure

You can use your own data by creating it within the following structure. Also you need to modify the app_params.py file with your data specifics as shown in the other examples.

./
└───data/
│   │
│   └───<dataset_name> (e.g. pets_2009)
│       │   └───train_im/(Image source files in .jpg format for training)
│   │   │    |    im (1).jpg
│   │   │    |    im (2).jpg
.
.
│   │   │    |    im (503).jpg
│       │   └───train_ann/ (Bounding Box Annotation Files for training)
│   │   │    |    im (1).xml
│   │   │    |    im (2).xml
.
.
│       │   └───val_im/(Image source files in .jpg format for validation)
│   │   │    |    im (504).jpg
│   │   │    |    im (505).jpg
.
.
│   │   │    |    im (600).jpg
│       │   └───val_ann/ (Bounding Box Annotation Files for validation)
│   │   │    |    im (504).xml
│   │   │    |    im (505).xml

Dependencies

Name	Version (min)
Python	3.7
Tensorflow	2.0.0
Numpy	1.18.5
Matplotlib	3.3.4
opencv-python	4.5.1
imagaug	0.4.0
larq	0.11.2

How to run

You run a training example as follows:

python det_keraspipeline.py -m resnet -d pets_2009 -s cosine -o adam -e 200 -lr 0.001

Argument Name	Description
-m, -model	Select backbone model from: [vgg,resnet,mobileV2,mobile,yolopeds,dronet,tinyyolov2]
-d,-dataset	Select dataset to use from: [kangaroo,ennfudan,pets_2009,raccoon,air_cars]
-e,-epochs	Number of epochs for training
-l,-lr	Initial learning rate
-o,-optimizer	Select optimizer from [adam,sgd]
-s,-scheduler	Select scheduler from [none, cosine, fix (reduce by 10 every 100 epochs)]
-w,-weights	Apply weights for imbalanced classes

• The training will output, live detection results for a random image per epoch, mAP and loss curves as shown below. All this are stored in the <det_output> folder:

• Models (both the complete architecture and weights file) are saved in the <saved_models> folder. In addition, everry 100 epochs a checkpoint is saved.

Backbone Models

The repository implements models of the following form:

Any of the backbone models can be selected. In addition, the detection layer can be configured according to the application needs.

• VGG16 (keras imagenet pretrained)
• Resnet50 (keras imagenet pretrained)
• MobileNetV2 (keras imagenet pretrained)
• MobileNetV1 (keras imagenet pretrained)
• YOLOpeds (custom)
• Dronet (custom)
• Tinyyolov2 (custom without imagenet weights)
• Any custom model that you want to create

Citation Information

This repository implements improved models from the following papers:

• Christos Kyrkou, “YOLOPeds: Efficient Single-Shot Pedestrian Detection for Smart Camera Applications”, IET Computer Vision, 2020, 14, (7), p. 417-425, DOI: 10.1049/iet-cvi.2019.0897 paper📜

• Christos Kyrkou, George Plastiras, Stylianos Venieris, Theocharis Theocharides, Christos-Savvas Bouganis, "DroNet: Efficient convolutional neural network detector for real-time UAV applications," 2018 Design, Automation & Test in Europe Conference & Exhibition (DATE), Dresden, Germany, pp. 967-972, March 2018. paper📜 , official DroNet repo 💻