Supplementary Materials of Visual Relationship Detection with Internal and External Linguistic Knowledge Distillation

Spatial features have been demonstrated to be helpful in visual tasks such as object detection, image retrieval, and semantic segmentation [13, 12, 9, 6, 1, 2, 7]. For visual relationship detection task, spatial features such as the relative location and size of two objects are informative for predicate prediction. For example, relative location is a discriminative feature for predicates “under” and “above” of the VRD training set shown in Figure 2(a) and 2(b). To explicitly model the spatial location and size of an object, we use the features from [10]: [ xmin W , ymin H , xmax W , ymax H , A