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.
The Images used (unlabeled and labeled ) are available here.
The images are collected manually and labeled using labelImg.
pyrcc5 -o libs/resources.py resources.qrc
python labelImg.py
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 |
- 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
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:
- 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