dariussingh / state-of-the-art-object-detection

State of the art object detection

State of the art object detection

Objective

The aim of this project is to use state of the art object detection algorithms for detection and classification of ticks.

Project Report

Dataset

Tick Dataset 1:

Contains 4 classes:

• Hyalomma Dromedarii Female: 27 images
• Hyalomma Dromedarii Male: 83 images
• Hyalomma Scupense Female: 17 images
• Hyalomma Scupense Male: 47 images

We see that the data is quite imbalanced and hence to tackle imbalance we use rotation as the Data Augmentation technique, we rotate the images with an angle of 90, 180 and 270 degrees only for the Hyalomma Scupense Female and Hyalomma Dromedarii Female classes as they are the minority class.

Tick Dataset 2:

Contains 11 classes:

• Hyalomma Dromedarii Female: 107 images
• Hyalomma Dromedarii Male: 211 images
• Hyalomma Scupense Female: 119 images
• Hyalomma Scupense Male: 147 images
• Hyalomma Impeltatum Female: 16 images
• Hyalomma Impeltatum Male: 46 images
• Hyalomma Marginatum Female: 24 images
• Hyalomma Marginatum Male: 60 images
• Other Female: 6 images
• Other Male: 16 images
• Hyalomma Excavatum: 22 images

We see that the data is quite imbalanced and hence we to tackle imbalance we use Data Augmentation. We rotate the images with a random angle between 0 and 270 degrees, we change the contrast of the image and we shift the RGB color of the image by a maximum of 30 units. We carry out Data augmentation for all classes except Hyalomma Dromedarii Female, Hyalomma Dromedarii Male, Hyalomma Scupense Female and Hyalomma Scupense Male as they are in majority

The dataset is not annotated hence we use makesense.ai to annotate the images in YOLO format.

Different models require the input data to be in different formats hence we convert the data into COCO and Pascal VOC format.

Models

We use the following models for object detection:

• DETR
• YOLOv5
• Faster RCNN
• Efficient Det
• SSD
• YOLOR
• Ensemble model of YOLOv5 and YOLOR

Code

The data_augmentation folder contains the data augmentaion notebooks to balance the various classes of the dataset.

Each model folder contains 4 notebooks:

• model trained on small dataset without augmentation
• model trained on small dataset with augmentation
• model trained on large dataset without augmentation
• model trained on large dataset with augmentation

Results

The results of the project are as follows: