CV Challenge

End-to-end computer vision project: dataset audit, augmentation, solid baselines (HOG+SVM/LogReg), a compact CNN, and robust evaluation (accuracy, macro-F1, confusion matrix).

Python NumPy Pandas OpenCV scikit-learn TensorFlow / Keras

View Repo Back to Projects

Project Overview

The Challenge

Build a reliable vision pipeline with clean data splits (no leakage), fair baselines, and a compact CNN. Emphasis on data quality, balanced augmentation, and explainable results.

The Solution

Implemented a reproducible workflow: audit/clean → stratified splits → augment → baselines (HOG+SVM/LogReg) → compact CNN with early stopping → evaluation via accuracy, macro-F1, confusion matrix and error analysis.

Project Details

Timeline

Started

2025

Duration

~6–8 weeks

Motivation

Hands-on computer vision challenge to practice clean ML engineering.

Technical Details

Software Used

VS Code, Jupyter, GitHub

Technologies

Python NumPy Pandas OpenCV scikit-learn TensorFlow/Keras Matplotlib

Skills Applied

Soft Skills

Problem solving & experimentation
Documentation & reproducibility
Time management
Communicating results

Technical Skills

Data auditing & augmentation
Classical ML baselines (HOG + SVM/LogReg)
CNN design & training
Evaluation (macro-F1, CM) & error analysis

Key Features

Clean Pipeline

Reproducible workflow with stratified splits and no leakage.

Baselines → CNN

HOG+SVM/LogReg baselines to benchmark a compact CNN.

Robust Evaluation

Macro-F1 and confusion matrix to understand failure modes.

Development Process

1. Data Prep

Audit, clean, split (stratified), and design augmentations.

2. Modeling

Baselines with HOG+SVM/LogReg → compact CNN with early stopping.

3. Evaluation

Macro-F1, confusion matrix, and error analysis → insights.