1 Additional Results 1 1.1 Synthetically generated partially annotated datasets 2 1.1.1 Knowledge-graph based partially annotated dataset generation

We use the two highest frequency ones which result in 776 label categories. Let us consider two datasets as shown in Figure 1. Let's say that the label Fine-grained mismatch problem: This problem occurs when a parent label ( e.g . We use the CIFAR100 [8] and MS COCO panoptic segmentation [7] datasets for this purpose. The third row has the similar thing, but for Dataset 2. Roughly Dataset 1 have 3x more data as the Dataset 2, with a total of 45k images across both.