Training Faster-RCNN on multiclass custom dataset

1. Introduction

Train object detector on multi-class custom dataset using Faster R-CCN in PyTorch.

This framework has the follow features:

• It is based on PyTorch framework
• It is designed to train on custom dataset
• It can train on multi-class dataset
• It automatically creates lables.txt file

2. Installations

2.1 Prerequisites

• Python 3
• Numpy
• PyTorch 1.8.1
• torchvision
• Pandas
• Opencv (cv2)

2.2 Code-Preparing

git clone https://github.com/harshatejas/pytorch_custom_object_detection.git

3. Dataset

This dataset contains images of playing cards.

The cards_dataset containds train folder, valication folder, train.csv and validation.csv

cards_datset/
 train/
   [xxx].jpg
   ...
 validation/
   [xxx].jpg
   ...
 train.csv
 validation.csv

4. Train

Modify Hyperparameters in train.py

train.py

5. Test

predict.py is designed to run predictions on the images in validation folder

Change the filename and saved_model in predict.py

predict.py

6. Predicted Images

Here are some sample output images predicted by saved_model/model