Containers in Montevideo: a Multi Source ImageDataset

This repo contains all code asosiated to the building process of the Clean-dirty containers in Montevideo (CDCM) dataset.

It was presented on October 19th, 2021 at the 50 JAIIO (Jornadas Argentinas de Informática). You can see the talk here, (it takes 20 minutes starting at 6:35:08, the link should leave you at this point).

There are also some demo videos available here. The videos were generated by processing each frame with the baseline models for the tasks purposed in the paper Containers in Montevideo: a Multi source Image Dataset. The logic goes as follows: first, containers are detected in each frame. Then, each container detected is cropped with a margin 20% larger than the detected container, and classified as clean or dirty. Finally, clean containers are marked in green and dirty containers in red.

In this repo

Since most of the code is experimental, it is provided as the following jupyter notebooks:

• download GSV images.ipynb: This notebook is used to download images from GSV based on the locations provided by the IM in the file data/Contenedores_domiciliarios.csv. You need to provide GSV API credentials to use it.
• evaluate garbage container classifier.ipynb: This notebook allows to train and evaluate a simple classifier to tell if a GSV image contains or not a container on it. It is trained over the Garbage Containers in Montevideo dataset.
• explore-containers-dataset.ipynb: This notebook is used to explore the CDCM dataset and to create the meta-data associated with it.
• container detection demo.ipynb: This notebook allows to process images or videos and to apply a detection model on them to get its containers, draw them on the image and save it. For example, it was used to create video demos like this.
• classification_baseline/classification-baseline.ipynb: this notebook contains the training and evaluation process for the baseline over the classification task.
• detection_baseline/explore metrics.ipynb: This notebook allows to explore the etrics stored as a result of the detection baseline model training, and it just show the detection_baseline/metrics.json file which contain those metrics. The training and evaluation for the detection baseline is performed in detection_baseline/run.py and it does require to download the dataset as in detection_baseline/get_data.sh.

Note: for obvious reasons, the datasets are not provided here, you need to download to the data folder.