Bottle Detection from Aerial Images

Abstract

Plastic pollution is a growing concern around the globe which possess long term environmental, health and economical threats. To minimize these threats computer artificial intelligence has stepped in with its domain computer vision to successfully identify the plastic waste in the wild. In this research, we propose a supervised learning object detection framework to find and localize waste bottles in the wild using UAV images dataset as plastic waste bottles are one of the top three most abundant plastic waste material but since bottles in UAV images are very small and sometimes transparent with complex backgrounds, it could be a very challenging task to correctly detect and localize such objects. For that reason, we have made use of ensemble methods since they can improve the object detection performance. In our implementation we have used voting strategy for dissembling the output of deep learning conventional neural networks (CNN) based object detection models since deep neural networks are fantastic at supervised learning and were able to outperform any corresponding model or technique. Best results were obtained by ensemble a strong single stage object detection model, RetinaNet with a powerful two stage object detection model, Faster RCNN with an AP value of 92%. Further, a detailed analysis of the dataset and benchmarks are presented in this research. This research also shows that choosing the right models for dissembling is crucial since in our testing we found that ensemble a weaker model with a strong one tends to decrease object detection performance, for that reason a detailed literature review was constructed and some existing models and techniques are presented in a brief comparison tabular form. Further, we have also showed the importance of data cleansing by the application of data preparation techniques, since going straight from data collection to model training leads to suboptimal results.