Method forward

models.py:895–932

(self, imgs, proj_mats, near_far, pad=0,  return_color=False, lindisp=False)

        return img_feat, in_masks

    def forward(self, imgs, proj_mats, near_far, pad=0,  return_color=False, lindisp=False):
        # imgs: (B, V, 3, H, W)
        # proj_mats: (B, V, 3, 4) from fine to coarse
        # init_depth_min, depth_interval: (B) or float
        # near_far (B, V, 2)

        B, V, _, H, W = imgs.shape

        imgs = imgs.reshape(B * V, 3, H, W)
        feats = self.feature(imgs)  # (B*V, 8, H, W), (B*V, 16, H//2, W//2), (B*V, 32, H//4, W//4)

        imgs = imgs.view(B, V, 3, H, W)


        feats_l = feats  # (B*V, C, h, w)

        feats_l = feats_l.view(B, V, *feats_l.shape[1:])  # (B, V, C, h, w)


        D = 128
        t_vals = torch.linspace(0., 1., steps=D, device=imgs.device, dtype=imgs.dtype)  # (B, D)
        near, far = near_far  # assume batch size==1
        if not lindisp:
            depth_values = near * (1.-t_vals) + far * (t_vals)
        else:
            depth_values = 1. / (1. / near * (1. - t_vals) + 1. / far * (t_vals))

        depth_values = depth_values.unsqueeze(0)
        # volume_feat, in_masks = self.build_volume_costvar(feats_l, proj_mats, depth_values, pad=pad)
        volume_feat, in_masks = self.build_volume_costvar_img(imgs, feats_l, proj_mats, depth_values, pad=pad)
        if return_color:
            feats_l = torch.cat((volume_feat[:,:V*3].view(B, V, 3, *volume_feat.shape[2:]),in_masks.unsqueeze(2)),dim=2)


        volume_feat = self.cost_reg_2(volume_feat)  # (B, 1, D, h, w)
        volume_feat = volume_feat.reshape(1,-1,*volume_feat.shape[2:])

        return volume_feat, feats_l, depth_values


class RefVolume(nn.Module):