Detection in Agricultural Contexts: Are We Close to Human Level?

摘要

We consider detection accuracy in agricultural contexts. Five challenging datasets were collected and benchmarked, with three recent networks tested. Based on an initial analysis showing the importance of image resolution, models were trained and tested with a multiple-resolution procedure. Detection results were compared to human performance, judged based on the consistency of multiple annotators. A quantitative analysis was made highlighting the role of object scale and occlusion as detection failure causes. Finally, novel detection accuracy metrics were suggested based on the needs of agriculture tasks, and used in detector performance evaluation.