An object-centric sensitivity analysis of deep learning based instance segmentation
In this study we establish a comprehensive baseline regarding the object-centric robustness of deep learning models for instance segmentation. Our approach is motivated by the work of Geirhos et al. (2019) on texture bias in CNNs. However, we do not compare against human performance but instead incorporate ideas from object-centric representation learning. In addition, we analyze and control the effect of strong stylization that can lead to disappearing objects. The result is a stylized and object-centric version of MS COCO on which we perform an extensive sensitivity analysis regarding visual feature corruptions. We evaluate a broad range of frameworks including Cascade and Mask R-CNN, Swin Transformer, YOLACT(++), DETR, SOTR and SOLOv2. We find that framework choice, data augmentation and dynamic architectures improve robustness whereas supervised and self supervised pre-training does surprisingly not. In summary we evaluate 63 models on 61 versions of COCO for a total of 3843 evaluations.
PDF Abstract
MS COCO
Stylized ImageNet
