Predicate classification and filtering

[dreichCSL]

Are there any details and/or performance measurements available for the Faster R-CNN model and the predicate classification in particular (object name, attributes and relationships)? How is the relation class predictor realized as part of FRCNN? In the paper you state using a regular CE (multi-task) loss, suggesting that each (ordered) pair of detected objects has exactly one relationship. Does every ordered pair have a relationship? For 100 objects that would be ~10k relationships for predicate input, which seems impractical to process. Or does relationship classification involve different different objects than the regular FRCNN output for object-id?

Some details would be appreciated. Thanks!

[Yunhao-Liu]

I have the same questions as you.

[liumc14]

Are there any details and/or performance measurements available for the Faster R-CNN model and the predicate classification in particular (object name, attributes and relationships)? How is the relation class predictor realized as part of FRCNN? In the paper you state using a regular CE (multi-task) loss, suggesting that each (ordered) pair of detected objects has exactly one relationship. Does every ordered pair have a relationship? For 100 objects that would be ~10k relationships for predicate input, which seems impractical to process. Or does relationship classification involve different different objects than the regular FRCNN output for object-id?

Some details would be appreciated. Thanks!

Do you know how the prediction of attribute and relational classes mentioned in the paper is handled? I also have this confusion, so please give me some guidance if you know，Thanks~

[dreichCSL]

Unfortunately, I don't know. I think, only the authors can provide this information.