Sticky Note Detector

This directory contains code implementation of custom trained SSD mobilebet v1 FPN model using my sticky-note images dataset i built to detect sticky-notes.

Requirements

Table of Contents

• Requirements
• Table of Contents
• Installations and Dependencies
  • Prepare The Dataset
  • Training and Evaluation
  • Testing
• Resources

Installations and Dependencies

Prepare The Dataset

The Images used (unlabeled and labeled ) are available here.
The images are collected manually and labeled using labelImg.

• labelImg Dependencies: python, PyQt5, lxml.
  
• Open cmd and go to the labelImg directory:
  

    pyrcc5 -o libs/resources.py resources.qrc

    python labelImg.py

Training and Evaluation

Library	version	Download Link
Tensorflow GPU	2.5.0	pip install tensorflow-gpu
NVIDIA CUDA Toolkit	11.4	official link
NVIDIA cuDNN	11.4	official link

Files to use (for tensorflow 2 and above)

• To start training the model

models\research\object_detection> python model_main_tf2.py /
                                  --pipeline_config_path=path/pipeline.config /
                                  --model_dir= path/to/model /
                                  --alsologtostderr

• To convert the save and export the model

models\research> python object_detection/exporter_main_v2.py /
                 --pipeline_config_path object_detection/path/pipeline.config /
                 --trained_checkpoint_dir=path/to/ckpt /
                 --output_directory=frozen_model

Testing

The model was able to recognize the sticky note with high confidence.
The testing procedure included both images that has sticky note/s and images does not have sticky note/s and it successfully detected all sticky note/s
A sample of the testing results:

Resources

• Dat Tran's raccoon_dataset : generate TFRecords files
• Tensorflow Object Detection API Documentation tutorial
• google, unsplash, pexels : used to building the sticky-note images dataset