A detail question about the model about pgcn

[Alvin-Zeng]

Thanks!
We indeed select the first row out of the N rows as the proposal feature to train our model. In the current version of P-GCN, for each proposal p, we first select the proposals that will affect p through multiple graph convolution layers. Then, we construct a proposal tree for this proposal relying on the contextual and surrounding rules. When the number of the neighbourhood is set to 10 and the number of GCN layer is set to 2, this tree will contain 1+10+10x10=111 nodes.

Please note that this proposal tree is equivalent to the proposal graph when we only take the node on top to train our PGCN model. We are now working on a graph-based version, and we will release it soon.

[ray342649093]

if you take a look at the pgcn_dataset file, you will then find that the first row of the features is the center prop features.(from the _video_centric_sampling, the _sample_adjacent_proposals, and the _sample_proposals_via_graph methods)

[HaiyiMei]

Thank you guys for the quick reply.