Monocular Depth Estimation and Data Analyzer

Intro

AI model training is just the beginning - accuracy is the key!

Then, I created a dedicated dataset analyzer to explore, compare and analyze results.

Data Analyzer

Main Challenge

• Reduce a relative error of 50% for humans in distance range between 100-125 [meters].

The Solution

• Develop a Data Analyzer tool to track on results.
• Data Augmentation such as Color Jittering, Rotations and more.
• Optimize the loss function to work on segmented pixels instead of a bounding box.

Data Analysis Tool

It is crucial to record how a model performes on a dataset, what are the downsides and track changes.
By using the tool - it is easy to find such inaccuracies in the ground truth / predictions, track them down, and consistenly improve the model.

Components

Confusion Matrix

Histogram

Statistics Table

Frameworks and Technologies

• Frontend: React, Redux, Material UI.
• Backend: C# Web API Core, Python, PyTorch, Numpy, OpenCV.