Deep Learning brought significant improvements both in terms of effectiveness and efficiency in various tasks related to robotic vision, including object detection and tracking among others. Object detection combines the tasks of classification and localization, aiming to find what objects are pictured in an image, as well as where in the image they are located, making it an extremely useful task in robotics applications. The most widely used object detection benchmarks include popular classes of objects, such as humans, cars, animals, etc., with PASCAL VOC and COCO depicting objects spanning 20 and 80 categories respectively. 

Despite improvements in these detectors on well known object detection benchmarks, deploying them on new applications highlights the domain adaptation problem. Domain adaptation refers to the process of alleviating the domain shift between a source and target domain, i.e., how to effectively deploy a detector trained on a source domain onto a target domain. Towards this purpose, it’s important to first evaluate the domain shift effect when deploying a detector on a new dataset, and to then use some form of knowledge transfer to effectively train a detector on the new dataset. 

Recently, AUTH in collaboration with AGI, have introduced the OpenDR Humans in Fields dataset, collected in the context of agricultural use cases using the Robotti robotic platform, designed for detection of humans in fields. Samples from the dataset can be seen in Figure 1, and the severity of the domain shift effect can be seen qualitatively in Figure 2, where a pretrained detector was deployed on the collected images. Specifically, after thorough evaluation of multiple pretrained detectors, the SSD model pretrained on the COCO dataset was chosen as the baseline, giving the best speed vs accuracy trade-off.