The effect of annotation quality on computer vision in efficient waste management

Deep learning-based computer vision models are often data-hungry, which has prompted the creation of larger datasets. The consensus within the field of computer vision is that larger datasets correspond to improved model performance. This is in line with the theory of generalisation which suggests that more complex models need larger datasets. However, the dataset quality is often not considered. Annotating datasets for fully supervised object detection and instance segmentation tasks necessitates substantial investments in time, effort, and other resources. In practice, the demand for large sample sizes often contributes to inaccuracies during the annotation process. This research seeks to comprehend and measure the influence of annotation quality on the performance of object detection and instance segmentation models. Furthermore, an investigation into whether the quantity or quality of annotations is of more importance is undertaken. In collaboration with an industry partner, we investigated the application of computer vision to automate and enhance efficiency within the waste management domain. The industry partner emphasised the critical importance of comprehending the trade-off between annotation quality and the associated investment, particularly given the pilot nature of this project. The findings from this work provide valuable insight into the effects of annotation uncertainty on mean average precision (mAP) performance. Finally, the results from our investigation into utilising computer vision to automate overfill bin status detection suggest that computer vision methodologies are both viable and effective for the task.