More detailed information about the project: http://opendata.letsdoitworld.org/#/ai
The model is meant to be used on google street view images and is taught to detect trash piles. If the model detects trash on any images, that do not include trash, then it means that it has not seen a similar object before in training dataset.
There are a lot of improvements to be made and a lot of new training images to be added to the project. Our intent is not to offend anyone or anything.
We have added the original dataset with some changes in the trash/dataset folder. It includes all the annotation json files (There are many, since they were done in different times and by different people).
Additionally there is a link to the images that were done by the LDIW volunteers during the cleanup days, so You are able to select and use the images yourself aswell here: https://drive.google.com/file/d/1X_ozEv5vF3bhg3FIIU6_5suBC7UdVVtA/view These images do not have annotations.
To try and test our model on your trash images:
- Download the latest h5 files from here: https://drive.google.com/drive/folders/1-ii6dHK3mUSY1mKfdYPPNZ18S7fEkl_o?usp=sharing
- Put the files into "weights" folder.
- Python environment requirements are described in requirements.txt
- Make sure you can use Jupyter Notebooks
Open the notebook: Detect_trash_on_images.ipynb If all the environment preferences match, you should be able to run the notebook.
Training of a image classificator is described here: https://github.com/matterport/Mask_RCNN
For image annotation we used VGG Image Annotator: http://www.robots.ox.ac.uk/~vgg/software/via/ Trash.py file is modified to understand the project-save- json files that come from VIA.
For training please add also coco weights from the drive folder to the weights folder.