The question about data augmentation about depth-from-motion HOT 5 CLOSED

batch_size = cur_features.shape[0]
num_depths = self.downsampled_depth.shape[-1]
h_out, w_out = cur_features.shape[-2:]
ws = torch.linspace(0, w_out - 1, w_out).cuda()
hs = torch.linspace(0, h_out - 1, h_out).cuda()
ds_3d, ys_3d, xs_3d = torch.meshgrid(self.downsampled_depth, hs, ws)
grid = torch.stack([xs_3d, ys_3d, ds_3d], dim=-1)
grid = grid[None].repeat(batch_size, 1, 1, 1, 1)
prev_cost_feats_list =[]
P2 = P2_ori.clone()
P2[:,0, :] = P2[:,0, :] // self.downsample_scale          
P2[:,1, :] = P2[:,1, :] // self.downsample_scale
for idx in range(batch_size):
    grid3d = points_img2cam(grid[idx].view(-1, 3), P2[idx][:3])
   
    pad_ones = grid3d.new_ones(grid3d.shape[0], 1)
    homo_grid3d = torch.cat([grid3d, pad_ones], dim=1)
   
    prev_grid3d = (homo_grid3d @ cur2prevs[idx].transpose(0, 1))[:, :3]
    prev_grid = points_cam2img(prev_grid3d, P2[idx])[:, :2]
    prev_grid = prev_grid.view(1, 1, -1, 2)
    prev_grid[..., 0] = prev_grid[..., 0] / (w_out - 1) * 2 - 1
    prev_grid[..., 1] = prev_grid[..., 1] / (h_out - 1) * 2 - 1

    prev_cost_feats = F.grid_sample(
        prev_features[idx:idx+1],
       prev_grid,
       mode='bilinear',
       padding_mode='zeros',
       align_corners=True)  # (B, C, 1, D*H_out*W_out)
       prev_cost_feats = prev_cost_feats.view(-1, num_depths, h_out,w_out)
       prev_cost_feats_list.append(prev_cost_feats)
prev_cost_feats_list=torch.stack(prev_cost_feats_list)  #(B, C, D, H_out, W_out)
cur_cost_feats_list= cur_features.unsqueeze(2).repeat(1,1,num_depths,1,1)
cost = (cur_cost_feats_list*prev_cost_feats_list).mean(dim=1)