All Case Studies
Computer Vision
VOC Segmentation App with Custom U-Net
OVERVIEW
The Project
A semantic-segmentation system that classifies every pixel in an image using a custom U-Net (ConvTranspose2d upsampling) on the Pascal VOC 2012 dataset, supporting all 21 classes — wrapped in a visual app for real-time exploration of predictions.
Objectives
Tools & Technologies
PyTorch LightningCustom U-Net (Conv2d / ConvTranspose2d)Weights & BiasesMatplotlib
METHODOLOGY
The Approach
1
Parse and convert RGB mask labels into integer masks; preprocess and normalize.
2
Build encoder–bottleneck–decoder U-Net; train with validation + metric logging.
3
Use ModelCheckpoint to save the top models by validation IoU.
4
Display batches and side-by-side image/mask/prediction grids in the app.
OUTCOME
Results & Learnings
Key Learnings
- RGB-mask → class decoding and batch normalization stabilize training.
- Balancing performance with real-time logging is a core tradeoff.
