Name: Semantic segmentation for application in very high resolution aerial images
Start: 2024-05-21T17:00:00+08:00
End: 2024-05-21T20:00:00+08:00
Location: DUG Technology

\*\*\* Note: Venues changed. Please read the rundown to find us in Kings Park Road, West Perth.\*\*\*

Rundown:
**5.00 - 5.45 pm** DUG Supercomputer Facility Tours and DUD Nomad Tours will be at **76 Kings Park Road** (for anyone wanting a tour before the presentation. A tour will take 20-30 minutes depending on questions.)
**6.00 pm Presentation will be at The Quest Hotel, 54 Kings Park Road** (150m walking toward the CBD)
\~7.00 pm After the presentation, networking at the DUG office, 76 Kings Park Road. (DUG kindly offer food and drinks during networking. DUG Supercomputer Facility Tours and DUD Nomad Tours can be arranged as well)

As a bonus, DUG will provide guided tours of Bruce: their immersion-cooled supercomputer facility. You will also get a tour of the new DUG Nomad - a mobile, modular, data-centre solution. [#DUGNOMAD](https://www.linkedin.com/feed/hashtag/?keywords=dugnomad)

Speaker bio:
Dr. Foivos Diakogiannis, Senior Research Scientist at Data61/CSIRO
presents a new modelling approach in semantic segmentation with application in very high resolution in aerial images and field boundaries.

Abstract:
In this talk, Dr Diakogiannis will present his latest work on semantic segmentation, introducing the SSG2 modelling framework (Semantic Segmentation Generation 2) as well as applications of in-field boundary detection using time series of input imagery. He will give emphasis on the attention mechanism and how this works in vision problems.

The core of SSG2 lies in its dual-encoder, single-decoder network, which is further augmented by a sequence model. This unique architecture accepts a target image along with a collection of support images. Unlike traditional approaches that rely on single, static images, SSG2 innovates by introducing a "temporal" dimension through a sequence of observables for each static input image. Within these sequence elements, the base model predicts the set intersection, union, and difference of labels from the dual-input images. This allows for more nuanced and accurate segmentation results.

The sequence model then takes the reins, synthesizing the final segmentation mask by aggregating these partial views and filtering out noise at each sequence step. The methodology is inspired by techniques in fields like astronomy and MRI, where multiple observations are utilized to enhance data quality.

It capitalizes on the strong correlation of true signals and the uncorrelated nature of noise across different sequence elements, offering a potential for statistical filtering and thereby reducing error rates.

\*\*\*\*\*\*\*\*\*\*\*\*\*\*\*\*\*\*\*\*\*\*\*\*\*\*\*\*\*\*\*\*\*\*\*\*\*\*\*\*\*\*\*\*\*\*\*\*\*\*\*\*\*\*\*\*\*\*\*\*\*\*\*\*\*\*\*\*\*\*\*\*\*

Sarada Lee

Sean Driver

Diamond Cutter

Perth Machine Learning Group

Technology

Data Analytics

Artificial Intelligence

Predictive Analytics

Natural Language Processing

Neural Networks

Deep Learning

Machine Learning

Artificial Intelligence Applications

Artificial Intelligence Programming

Python

Semantic segmentation for application in very high resolution aerial images 

DUG Technology

Share

Perth Machine Learning Group

Semantic segmentation for application in very high resolution aerial images

Perth Machine Learning Group

Details

Members are also interested in