Humans need not label more humans: Occlusion Copy & Paste for Occluded Human Instance Segmentation

Modern object detection and instance segmentation networks stumble whenpicking out humans in crowded or highly occluded scenes. Yet, these are oftenscenarios where we require our detectors to work well. Many works haveapproached this problem with model-centric improvements. While they have beenshown to work to some extent, these supervised methods still need sufficientrelevant examples (i.e. occluded humans) during training for the improvementsto be maximised. In our work, we propose a simple yet effective data-centricapproach, Occlusion Copy & Paste, to introduce occluded examples to modelsduring training - we tailor the general copy & paste augmentation approach totackle the difficult problem of same-class occlusion. It improves instancesegmentation performance on occluded scenarios for "free" just by leveraging onexisting large-scale datasets, without additional data or manual labellingneeded. In a principled study, we show whether various proposed add-ons to thecopy & paste augmentation indeed contribute to better performance. OurOcclusion Copy & Paste augmentation is easily interoperable with any models: bysimply applying it to a recent generic instance segmentation model withoutexplicit model architectural design to tackle occlusion, we achievestate-of-the-art instance segmentation performance on the very challengingOCHuman dataset. Source code is available athttps://github.com/levan92/occlusion-copy-paste.