Custom-Dataset-Training-with--DE⫶TR

Custom Dataset Training pipeline using Pytorch and Meta's object detection model DE⫶TR.

DE⫶TR Detection Transformer

DETR (DEtection TRansformer) is object detection model developed by Meta AI.

Paper and repository here.

Dataset

VisDrone2019 Dataset is an object detection and tracking dataset that consist of around 10k images. To read more...

Images from dataset

1	2

Project Structure

 Custom Dataset Training with DETR
 ├───detr
 │   ├───datasets
 │   ├───models 
 │   └───util     
 ├───VisDrone2019
 │   ├───Train
 │   │   ├───annotations
 │   │   └───images
 │   └───Validation
 │       ├───annotations
 │       └───images
 └───weights

Training Pipeline

• First you need an custom dataset class or you just can use my VisDroneDatasetClass.
• Secondly just adjust the transformation functions on utils.py
• Finally add those changes you made in other files onto run.py file and run.

What's next?

[ ] Add weights that trained on VisDrone2019.
[ ] Add a demonstration of Model.