It is werid to use GT mask for depth_pre in compute_depth_losses

[lmomoy]

def compute_depth_losses(self, inputs, outputs, losses, accumulate=False):
    """Compute depth metrics, to allow monitoring during training

    This isn't particularly accurate as it averages over the entire batch,
    so is only used to give an indication of validation performance
    """
    depth_pred = outputs[("depth", 0, 0)]
    gt_height, gt_width = inputs['depth_gt'].shape[2:]
    depth_pred = torch.clamp(F.interpolate(
        depth_pred, [gt_height, gt_width], mode="bilinear", align_corners=False), 1e-3, 80)
    depth_pred = depth_pred.detach()

    depth_gt = inputs["depth_gt"]
    mask = depth_gt > 0

    # garg/eigen crop
    crop_mask = torch.zeros_like(mask)
    crop_mask[:, :, 153:371, 44:1197] = 1
    mask = mask * crop_mask

    depth_gt = depth_gt[mask]
    depth_pred = depth_pred[mask]
    depth_pred *= torch.median(depth_gt) / torch.median(depth_pred)

    depth_pred = torch.clamp(depth_pred, min=1e-3, max=80)

    depth_errors = compute_depth_errors(depth_gt, depth_pred)

    for i, metric in enumerate(self.depth_metric_names):
        if accumulate:
            losses[metric] += np.array(depth_errors[i].cpu())
        else:
            losses[metric] = np.array(depth_errors[i].cpu())

If this operation will introduce GT information for prediction?

[daniyar-niantic]

Hi @lmomoy

As the comment to the function indicates, this is not used for training (no backpropogation on these metrics) but to monitor how well the model is doing during training.