Computer Vision (TensorFlow Lite)

program that can find patterns or make decisions from a previously unseen dataset.

Main steps

1 - Video/Roboflow: Gather images for dataset

2 - Colab: Train Custom Model

3 - Export Model

4 - Set up Pi

5 - Run on Pi

1 - Video/Roboflow: Gather images for dataset

Logo	Description
	We can film a video of the item we want to detect and then upload to Roboflow. On the Roboflow website, we can extract frames every X seconds to create a .jpg file each interval. Then we can begin labeling each image on Roboflow.

Roboflow is a website where you upload images, label them, and export a zip file (images & .xml files).

Useful to share dataset and have 3 people label at the same time

1.1 - Unlabeled Images to Labeled Images

Unlabled	Labeled (Desired Outcome)

2 - Colab: Train Custom Model

WARNING: This Colab began having issues in May 2023.

Logo	Description
	-The following website uses Python -here we set up libraries - collections of pre-written code that you can use to perform specific tasks -Prepare the dataset -Train the model -Evaluate the model performance with validation data -Export the model as .tflite file

Open up this pre-made online coding website called Google Colab

3 - Export Model

We can export the custom model after training as a .tflite file which will be uploaded to the Raspberry Pi and then can be used for Object Detection. Demo here

4 - Set up Pi

Original Tensorflow Video - 2 yrs ago	November 2023 - Recent Video
YouTube Link	YouTube Link
Pi Versions numpy >= 1.20.0 opencv-python~=4.8.0 tflite-support>=0.4.2 protobuf>=3.18.0,<4 NOTE:: - python 3.7.3	Colab Versions numpy 1.23 opencv-python-4.9.0.80 tflite-support==0.4.2 pip install protobuf==3.20.* NOTE:: - Python 3.9.18 - pip install ipykernel - conda create -n myenv python=3.9 -y - pip 23.3.1 - pycocotools-2.0.7

---

Installation

Original Tensorflow Video - 2 yrs ago	November 2023 - Recent Video
YouTube Link	YouTube Link
NOTE: (2:54) -click top left -preferences -raspberry pi configuration -tabs: interfaces -enable camera -reboot	NOTE: -click top left -preferences -raspberry pi configuration -tabs: interfaces -enable camera -reboot
(3:26) sudo apt-get update python3 --version NOTE:: python 3.7.3 python3 -m pip install --user --upgrade pip python3 -m pip install --user virtualenv python -m venv ~/tflite source ~/tflite/bin/activate git clone https://github.com/tensorflow/examples.git cd examples/lite/examples/object_detection/raspberry_pi sh setup.sh python detect.py	sudo apt-get update sudo apt-get install python3-tflite-runtime git clone https://github.com/tensorflow/examples.git --depth 1 cd example/object_detection/raspberrypi/ NOTE: change opencv version nano requirements.txt opencv-python~=4.8.0 #from tflite official pg bash setup.sh pip install protobuf==3.20.* python -m pip install --upgrade tflite-support==0.4.2 #didnt download python -m pip install --upgrade tflite-runtime==2.11.0 #didnt download pip install numpy==1.23 (33:16) opens detect.py-in thonny and then edits custom model and then runs it to detect #Python 3.10.12

5 - Run on Pi

Original Tensorflow Video - 2 yrs ago
source tflite/bin/activate
cd examples/lite/examples/object_detection/raspberry_pi/
cp ~/Downloads/android.tflite .
python detect.py
python detect.py --model android.tflite

Sources

[1] Tensorflow Videob (The Colab they have doesn't work with newer Python versions)