Promoting Usage of Deep Learning Object Detection in Ecology by Improving Performance and Accessibility

The inability of object detectors to generalise to domains beyond those included in labelled training data is limited when the training data has high intra-dataset similarity. This dataset aims to address this by providing data characterised by high intra-dataset variability. Highly variable images were scraped from FlickR and iNaturalist using python scripts available at https://github.com/ashep29/infusion for the following animals: Sus scrofa, striped hyena, and rhinoceros. These were supplemented with location specific camera trap images from WCS Camera Traps (WCS_striped_hyena and WCS_rhino), Snapshot Serengeti (SS_striped_hyena and SS_rhino), Missouri Camera Traps (EU_pig) and North American Camera Trap Images (NA_pig) which are publicly available on www.lila.science. The high intra-dataset variability of these subsets was ensured by removing all images with an SSIM score greater than 0.8 (where 1.0 represents identical images). All these images were then annotated in PASCAL VOC format with bounding boxes to allow for object detector training.