TrashAsistant: A kivy App, which uses Deep Neural Networks...
summary
Prototype mobile app built with kivy app that classifies trash photos using VGG16 transfer learning and a custom top, trained on about 4,000 images. Two labeling schemes (4 classes, 8 classes) produce best reported test accuracy near ~85% for four labels, lower for eight; glass/plastic confusions and reflections remain common. Training runs in Google Colab with GPU/TPU, model exported as .h5 and loaded client-side, demo is early alpha and crashes intermittently.
AI-written summary · about
chapters (14)
- 0:00TrashAsistant project intro
- 1:13Motivation for trash segregation app
- 3:24EU regulations and Gdańsk context
- 5:37App concept and data collection
- 7:25Transfer learning for image classification
- 9:21Dataset preprocessing steps
- 11:53Model training with Keras and Colab
- 13:44VGG16 customization details
- 15:594-class vs 8-class model comparison
- 18:08Kivy framework for app development
- 21:42TrashAsistant prototype demo
- 24:37Model accuracy challenges
- 25:49Future improvements and roadmap
- 26:55Q&A and municipality plans
related talks
31mFinding trash in waste
34mMachine learning for front-end developers
29mLabeling Tools are Great, but What About Quality Checks?
58mDeep Learning in Real Life
37mJavaScript Machine Learning from Magic to Science
28mWhy I didn't use deep learning for my image recognition
51mA Practical Introduction to Convnets for Image Classification
35mComputer vision: British Clay