expand_less
From the interview with Prof. Ngo, we realised that the main cause of plastic pollution is that plastic is too convenient for the general public, and this leads to an excessive usage of plastics.
He pointed out that in order to tackle this problem, we must organize more promotion events for citizens to arouse the awareness of the new generation. In particular, educating our citizens about the environment at a young age can help stimulate correct environmental perceptions. As for how to raise environmental awareness of adults, we should collaborate with NGOs and transfer data to different departments.
Moreover, using social media to promote our project and share the severity of plastic pollution can also be a useful method. Youtubers and idols are mostly generation-Z, and they may be willing to help us share information about the environment.
We had made use of help from our schoolmates to tag plastic bottles in our photos. After tagging the photos, they will understand the real situation of plastic pollution and be more aware of this problem. Furthermore, this process of plastic tagging can convey the message that the process of removing plastic bottles is very complicated, and that cutting-edge technology, such as drones and neural network are needed to efficiently remove plastic bottles from the environment. This is a way of raising awareness about plastic pollution.
Our Human Practice team has managed the Instagram account of our team, presenting some of our tagging results and AI data. We hope that the general public will be more attentive towards plastic pollution and that they will be more willing to tackle this problem together through our promotion.
According to Prof. Chan’s opinion, our overall plan of creating a model that detects plastic bottles is good, but we must set clear targets and procedures. He suggested that there could be other external factors affecting the number of plastic bottles on beaches. Possible natural factors include tidal variations, and possible human factors include decreased number of beachgoers during the pandemic of COVID-19). He also suggested that the number of plastics might fluctuate throughout different times of the day, and depend on the amount of human activity on the beach.
Prof. Chan also proposed focusing on several beaches in Hong Kong for investigation, and noted that the guidelines for tagging plastic bottles must be clear, since besides the errors made by the Deep Learning Model, human errors are also significant.
Responding to our problem of the number of photos we should take to achieve the maximum model performance, he suggested that we carry out power estimation using random subsampling through trial and error for model training, and through different batches of different sizes to determine the maximum accuracy of the model.
Our team has visited Lantau Island, Cheung Chau and Lei Yu Mun to conduct our plastic photography so that we can get a grasp of the plastic problem in general in Hong Kong. Moreover, we have made a plastic tagging guideline document and also filmed a video explaining the steps of plastic tagging for our student volunteers.
Moreover, we have implemented power estimation by dividing training data to 25%, 50%, 75%, and 100% to our model to determine the maximum efficiency of our model.
For the Drone & AI team, Mr. Sidhant taught us about some basic concepts of machine learning and Convolutional Neural Networks, as well as some tips on model training.
In response to our PET bottle detection work, Mr. Sidhant’s team suggested that we use either one of three algorithms: YOLO, Faster-RCNN or Mask-RCNN. They also recommended us to use Detectron, an object detection codebase developed by Facebook, as a framework.
In addition, he pointed out some droning locations including Lantau island and Lamma island, which are some of the worst polluted places in Hong Kong.
As our Drone & AI team is aiming to detect PET bottles in non-real time, we have chosen the Mask-RCNN implemented from Detectron, as suggested by Mr. Sidhant. We have divided our dataset to two sections: Train and validation, for both training and performance testing in a ratio of 8:2.
Also, we configured the training iterations to around 1000 as suggested until the accuracy and loss of the model have reached maximum and minimum respectively, to determine the number of iterations for the model to achieve maximum performance. Nevertheless, they also lent their image annotation platform - CVAT for us to assist our data preparation work.
