Vehicle Object Detection
Aerial View Point

In this project I develop a non-axis aligned 2D bounding box object detection algorithm based on the EfficientDet architecture (currently state of the art in 2020). See below the architecture diagram for efficientDet and it's performance against competitor models.

To begin with I hand labelled data with rotated bounding boxes using this tool.

I then created an EfficientDet model and loaded weights pre-trained on Imagenet. I used this implementation of the EfficientDet architecture.

See below the tensorboard loss curve for training.

See below the output video on data never seen before by the algorithm during training: