# Capstone Meeting - 20/09/2020

* Assessment review (Poster, Endeavour Presentation Video):
* due - first week of the october;
* [graded] poster - Friday, 9th October; [ref] https://endeavour.unimelb.edu.au/students/workshops
* [not graded] endeavour presentation; - Friday, 9th October
* Finalise Objectives for our project, and  Specifications of our System:
* Project updates
* Setting what needs to be done and what is out of scope
* Documentation

## Assestment Review
* Poster [Link](https://www.overleaf.com/project/5f4cecfe2b40f00001f23535)
* due - first week of the october;
* [graded] poster - Friday, 9th October; [ref] https://endeavour.unimelb.edu.au/students/workshops
* [not graded] endeavour presentation; - Friday, 9th October

## Objectives & Specifications

#### Project Objectives:
* Translate emergency auslan signs in real-time;
* Bring forth the awareness of the auslan community towards the public;
* [KIV] - building an application on top of real-time pose estimation; (skeleton based);
* [KIV] Integrate our knowledge in DSP and and machine learning to build a real-world application?
* [KIV] Showcasing signal processing techniques and machine learning in building a real-world application?

#### System Specifications
* System successfully shows a proof of concept of recognising signs using human pose estimation.
* Technical Specifications/Metrics
* Accuracy:
* [Notes] - Applying it on validation/test set, real world
* Quantity - How many people successfully picked up? generalization;
* Quality - How well does it pick up?
* Speed:
* Factors
* Internet Connection Quality - CISCO VPN, WiFi, User's Internet.
* Computation Limit - GPU Specs on Cloud/Desktop's GPU power to do inferencing.
* [Notes]:
* Let's set an average internet connection rate to define our system speed based on **INTERNET CONNECTION**
* We can test out system on a few different systems to see how well it performs to infer signs.

* Number of signs
* Able to recognise up to 5 UNIQUE Auslan signs continuously.
* As you increase sign classes, training data increases
* [flag - report] Highlight limitation
* isolated signs

## Project Updates

### Frame reduction
* Reducing sampling rate from 75 to 35.
* Automated into datapipeline.
* Steps:
1) Cut from 75 to 70 first.
2) Exclude frames that have zero keypoints.
3) if you still have 70 frames;
* Get odd and even sample frames.
* Doubled dataset using both odd and even frames as data.

### Sign Updates
* Hospital, Ambulance, Pain, Thumbs Up, Help.
* Increase variability by recording different data.

### Application Updates
* Test by ONLY reducing window width.
* Re-deploy using WebRTC

## Moving Forward
* Evaluation:
* Read up/Research on valid metrics that we can use in our project to define specs (accuracy); (Yick and Tsz Kiu)
* speed (Matthew and Yick)
* Poster & Presentation:
* Poster (Tsz Kiu and Matthew)
* Presentation Slides - Matthew
* Meet up middle of next week to discuss.
* research limitation;
* read up on exploding gradient vs vanishing gradient;

* [yick]
* batch generator;
* hyperparameter with spark
* [tune] number of hidden units:
* lstm layer;
* Dense Layer Numbers
* [tune] Dropout rate:
* discretize to fix within [0.1, 0.9];
* Learning Rate
* Default value is quite ok
* Decay Rate
* default value is quite ok;
* Optimizer - Adam
* augmentation,
* compensation;
* tweak variability;
* [Matt]
* Port to WebRTC to test speed with old system.
* Test out new model.