aka-vm / Hello-CV

This repository performs Computer-Vision tasks on multiple Image Datasets using CNN based Networks.

Hello CV

Not just a hello-world Repository with entery-level code!

In this repo I have worked to develop various Computer Vision taskes. The code uses both Pytorch & Tensorflow.

TODOs for running the code

1. Create paths.py file (use paths.py.template as a template)
2. Download the datasets.

Datasets

All the used datasets well-known around the deeplearning community. They include -

• MNIST - Digit Recognizer: A dataset of handwritten digits.
• CIFAR-10: This dataset contains 32x32 color images of 10 classes of objects.
• Stanford Dogs Dataset: Contains images of 120 breeds of dogs. Each image has different pixel values and ratio.

Data Preprocessing

• For large datasets I have used Image Iterator, This uses less memory.

• Image augmentation is also used to augment the images.

Image Classification Results

S. No	Dataset	Best Accuracy	Best Model	Real-Life Test
1	MNIST(Digit Recognizer)	99.3%	Simple Sequential CNN
2	CIFAR-10	89.94%	VGG-Like Model	5 images
3	Stanford Dogs	84.3%	InceptionResNetV2	4 images

Other Computer Vision Tasks

S. No	Application Name	Refered Literature/Implimentation	Implimented Using	Metric	Score	Visuals
1	Neural Style Transfer	A Neural Algorithm of Artistic Style, NST With Two Style	Pytorch
2	Dog Breed Detection (YOLOv8)	Joseph Chet's Publications, YOLOv8 Implimentation	Pytorch, Ultralytics	mAP50-95	0.79

---

Check the branches. I have created a new branch everytime I added a new type of model.

NOTE: For Some reason different hardware give different results, I used two Machines, Macbook Air M1(8gb) and Intel i7 11700k | RTX3070. RTX3070 machine gave better results with a good margin. I even used Kaggle and Jarvis Labsto train some of the models.