CUT TO:
INT. PROJECT ARCHIVE — STORYBOARD ROOM
The USER opens Image Segmentation Model Comparison.
DIVAKAR DESSAI
Let's run through the shots.
CASE FILE / COMP9517 / Computer Vision / Machine Learning
Image Segmentation Model ComparisonClassical ML vs Deep Segmentation Networks
Compared classical machine-learning and deep-learning approaches to image segmentation, including K-Means, Random Forest, SVM, FCN, U-Net and ResUNet.
01–02 / OPENING SEQUENCE
Establishing the world
01 / Establishing Shot
The Problem
Different image-segmentation techniques make very different assumptions about how visual structure should be represented, making direct experimentation important when selecting an approach.
02 / Wide Shot
The Context
The work formed part of a COMP9517 computer-vision group project comparing classical machine-learning methods with modern deep segmentation architectures.
03 / CHARACTER NOTE
DIVAKAR'S ROLE
Subject
Divakar Dessai
Production
Image Segmentation Model Comparison
Take
03 / Role
Role notes
DIVAKAR DESSAI
I contributed to experimentation, implementation and comparison across several image-analysis and segmentation approaches.
04 / CLOSE-UP
THE ENGINEERING CHALLENGE
The methods ranged from unsupervised clustering to fully convolutional deep networks, requiring different data preparation, training and evaluation workflows.
05 / TRACKING SHOT
THE APPROACH
A plan emerges.
The project implemented K-Means, Random Forest, SVM, FCN, U-Net and ResUNet pipelines and compared their behaviour and modelling assumptions.
06 / INSERT SHOTS
KEY FEATURES
The system takes shape.
07 / DIRECTOR'S NOTES
DESIGN DECISIONS
Things we decided along the way
01
Compared fundamentally different modelling families rather than assuming a neural model was always preferable.
02
Used segmentation-specific encoder-decoder architectures for dense prediction.
03
Explored skip connections and residual learning through U-Net and ResUNet.
design decisions
somewhere mid-build
08 / RETAKES
WHAT WENT WRONG
Naturally, not everything cooperates.
09 / FINAL SHOT
THE OUTCOME
ProblemBuildOutcome
01
Implemented and compared six computer-vision approaches.
02
Developed practical understanding of semantic segmentation architectures.
03
Gained experience with deep-learning training and model evaluation.
10 / PRODUCTION NOTES
TECH STACK
The tools behind the scenes.
PythonJupyterK-MeansRandom ForestSVMFCNU-NetResUNet
11 / BEHIND THE SCENES
LINKS
FADE OUT.
USER closes the file.
One project down. A few more stories left.
DIVAKAR DESSAI
Pick another?