RecoMatrix

AI Enigma Contest task on real time complex object detection using computer vision and machine learning in python. The script detects most of the real world objects, human and pets in real time.

Algorithms planned to be implemented

Click to expand

• Faster R-CNN
• R-FCN
• SSD

Project Structure

COCOFasterR-CNN

complexObjectDetection
|--- output
|--- scripts
|--- src
|  |--- coco.names
|  |--- detector.py
|  |--- main.py
|--- test
|--- env.yml

Directory Details

• output: This directory is for keeping the detected new images. Used by the createBoundingBox method in detector.py file
• trained_model: After testing against Faster and Mask R-CNN, finally,Faster R-CNN which is generated by the downloader from TF Object Detection Model Zoo
• scripts: A directory created & defined by the developers to keep useful scripts for making the developement of this project as seamless as possible
• src: Holds the python codes
• test: It is used to store the existing and real time complex image and videos which are used by the main.py file for training the model.

Environment Setup

The project environment is managed by conda . Details are given below:

• OS Platform: Linux 6.5.5-arch1-1
• Python Version: Python 3.9.0
• PIP Version: pip 23.2.1
• Conda Version: conda 23.7.4
• Tensorflow Version: tensorflow-gpu 2.6.0
• Cuda Toolkit Version: cudatoolkit 11.2.2
• CUDNN Version: 8.1

The above mentioned python packages version matching is an important part. It must be noticed that if the libraries versions are matching or not. There are two ways to make the developement environment ready:

Automatic - Using the YML file

I have created an env.yml file using conda so that it becomes easier to install the developement libraries in one command. Use this command to install all the required libraries with specific versions.

Note that conda must be installed first

conda env create -f env.yml

Manual

Every packages can be installed manually. The advantage of installing manually is that we get more control and it gets easier to solve errors and also find errors.

• Create the virtual environment first

conda env create --name tf_gpu python 3.9
conda activate tf_gpu

• Install the libraries

conda install cudatoolkit=11.2 cudnn=8.1 -c=conda-forge
pip install tensorflow-gpu==2.6
pip install opencv-python

Running the project

We have made running the project much easier

• Activate the conda virtual environment first

conda activate tf_gpu

• Go to the src directory

cd src

• Run the main file

python main.py

Command Line Arguments

We have added some CLA to out project so that we don't need to change in code when we want to run in different modes of our project. For example, currently we have a --camera argument which will activate the live detection using the webcam

python main.py --camera

Currently supported CLA are:

--help      Print the built in help menu for this program
--camera    Enable the live detection using camera

Performance Analysis

Meta Architecture (Accuracy and Speed)

Meta Architecture (Object Resolution)

Meta Architecture (Feature Extractor)

Architecture (Proposals Count)

References

• COCO Dataset
• TF Object Detection Model Zoo

Authors