search

lmur98 / AFFttention

Official repository of "AFF-ttention! Affordances and Attention models for Short-Term Object Interaction Anticipation"
0 stars 0 forks source link

Error with propagating global features to ROI head #2

Open wisdomikezogwo opened 2 months ago

wisdomikezogwo commented 2 months ago

Hi, wanted to know how you propagate the global feature which i presume is of size B * dim to the ROIhead using this line without throwing an error with the RPNHead which uses all the features from the backbone and expects all features (which is the multiscale features) to be of size BS, C, W, H?

lmur98 commented 2 months ago

The RPN head ignores those features and only work with the multi-scale features OrderedDict([("0", patch_2D_fused), ("1", patch_2D_fused), ("2", patch_2D_fused), ("3", patch_2D_fused), ("pool", patch_2D_fused)])

This is very similar to the StillFast head,