Skip to content

Fork of a project I have worked on

Notifications You must be signed in to change notification settings

doughnut923/HK_GTC

 
 

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

49 Commits
 
 
 
 
 
 
 
 
 
 

Repository files navigation

HK_GTC - Detection Of PET Bottles

As we all know, plastic pollution is a severe problem to our ecosystem. In respond to this problem, our team had created a plastic detecting model. In this Github Repository, you will know how to use the model, and train the model with your own dataset.

Dependenies

  • Local Linux environment or Google Colab Notebook
  • Detectron2
  • Jupyter Notebook
  • pytorch 1.8
  • torchvision
  • OpenCV
  • Numpy

Training

You can either train using our Colab Notebook or in a local linux enviornment.

Local Training / Using the Model

Clone the repository and open colab.ipynb for model training and usage. Details of using the codes will be included in the notebooks.

Training Results

We've trained our photos with different backbones and sample size. Backbones can be found in detectron2/MODEL_ZOO. The samples are selected randomly from the training dataset. The configurations used for all training are as follow:

  • Backbone used: X101-FPN3x, R101-FPN3x, R50-FPN3x
  • Batch size per step: 2
  • Iterations: 1000
  • Train image percentages used: 100%, 75%, 50%, 25% of training data

Backbone test

Backbone AP AP50 AP75 APs APm APl model
R50-FPN3x 51.1 72.2 60.6 16.7 44.8 68.3 model
R101-FPN3x 52.3 74.8 59.9 19.0 45.7 69.0 model
X101-FPN3x 54.8 77.8 61.4 10.9 48.8 72.5 model

Power estimation test

(Models here are all using R50-FPN backbone)

Train Images Percentage AP AP50 AP75 APs APm APl
100 51.1 72.2 60.6 16.7 44.8 68.3
75 49.6 70.0 58.2 12.0 44.0 66.7
50 48.6 70.4 58.0 12.9 43.0 64.0
25 48.0 68.4 57.5 11.9 43.4 63.0

About

Fork of a project I have worked on

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages

  • Jupyter Notebook 100.0%