How Much Augmentation is Needed for Single-Class Segmentation?
Loganbirdy · reddit · 2026-07-14
The author is training a single-class segmentation model to identify large rectangular artworks photographed on floors, aiming for higher boundary accuracy.
Data & Scenarios
- About 3,000 high-quality originally annotated images
- Sourced from 6 photographers
- Varying shooting conditions:
- roll / pitch / yaw
- camera distance
- frame coverage, center offset
- orientation, perspective
- lighting
- Images captured on flagship iPhones
Problem
They want to use on-the-fly augmentation to simulate real-world handheld variations, reducing manual leveling work for designers. They ask:
- Is generating 100 augmented sets per image helpful, or overkill?
- Should the augmentation strategy lean towards:
- Individual transforms
- Stacked transforms
- Controlled mixing
Goals
- Maximum segmentation accuracy, especially at object boundaries
- Training for about 300 epochs
- Keeping validation and test sets unaugmented
More from Research
- New paper defines self-state attacks, showing OS defenses leave four agent-memory cases indistinguishable — Justgototheeffinmoon · 2026-07-22
- Krea 2 users recommend a two-pass Clownshark sampler setup for sharper image details — listopalafoto · 2026-07-22
- Animation shows how an MLP’s first-layer weights change while learning MNIST — CatAstro_Piyush · 2026-07-22
- Project APE finds verifier reliability drops when papers contain multiple errors — soumitrashukla9 · 2026-07-22
- Project APE says verifier costs fell about 90x in a year as Chinese open models lead — soumitrashukla9 · 2026-07-22
- OpenAI-linked paper says capability RL can make models more reward-seeking — MariusHobbhahn · 2026-07-22