some question about using the features after RCNN

[sunbear616]

hello, Thanks for your code, actually, I want to try a new way to detect the visual relationship, I saw the paper there is a Fusion layer. I didn't find this part, maybe in VRD_model? but I want to ask, what you get from the fusion layer, the score for object and relationship? and if I want train a new network to make VR detection, should I use the feature before Fusion layer or after? I am totally fresh man in Fast-Rcnn, and I saw some paper, I think everyone use the different way, but in this paper, actually I really want to see how Relationship Instance looks like. so which part of code I can get the this Relationship instance？ Thanks a lot

[GriffinLiang]

Fusion layer is a fully connected layer which takes the concatenation of visual, spatial and semantic cue as the input. https://github.com/GriffinLiang/vrd-dsr/blob/c598c0139feb6ccd137feeceeae898c2cbfd68f1/lib/nets/Vrd_Model.py#L75

[sunbear616]

Fusion layer is a fully connected layer which takes the concatenation of visual, spatial and semantic cue as the input.

vrd-dsr/lib/nets/Vrd_Model.py

Line 75 in c598c01

self.fc_fusion = FC(n_fusion, 256)

hey thank you for your replying, and may I ask you, how can you prepare the VRD or VG dataset, I found in tools folder there is a train.py, But when I train it , it shows the error about my dataset, so Did preprocessed dataset like in faster rcnn? right?