search

WXinlong / SOLO

SOLO and SOLOv2 for instance segmentation, ECCV 2020 & NeurIPS 2020.
Other
1.69k stars 307 forks source link

Questions about solo head #124

Open talebolano opened 3 years ago

talebolano commented 3 years ago

This is an excellent model, but I have questions about the realization of the solo head. In the source code, all prediction branches share the same weight. 

        # cate branch
        for i, cate_layer in enumerate(self.cate_convs):
            if i == self.cate_down_pos:
                seg_num_grid = self.seg_num_grids[idx]
                cate_feat = F.interpolate(cate_feat, size=seg_num_grid, mode='bilinear')
            cate_feat = cate_layer(cate_feat)

        cate_pred = self.solo_cate(cate_feat)

Is this just for reducing the parameter size, or will it improve the prediction accuracy?

zzzzzz0407 commented 3 years ago

@talebolano both, actually almost all the dense detectors, \eg, RetinaNet, FCOS, and so on, share the weights in their heads.