metrics.py

# Copyright (c) Meta Platforms, Inc. and affiliates.

import numpy as np


class MADEComputer:
    def __init__(self, is_valid_pixel_mask: np.ndarray, log_depth_gt: np.ndarray):
        assert is_valid_pixel_mask.ndim == 2 and is_valid_pixel_mask.dtype == bool
        assert log_depth_gt.shape == is_valid_pixel_mask.shape
        self._H, self._W = is_valid_pixel_mask.shape
        self._is_valid_pixel_mask = is_valid_pixel_mask
        self._depth_gt = np.exp(log_depth_gt)

    def compute_curr_made(self, log_z: np.ndarray, return_scale: bool = False):
        # Based on https://github.com/xucao-42/bilateral_normal_integration/blob/
        # e819fc11742c1316bcc649d078848096b6d77e85/evaluation_diligent.py#L41-L44.
        assert log_z.shape == (self._H, self._W)
        depth_map = np.exp(log_z)
        scale = np.nanmedian(self._depth_gt / depth_map)
        scaled_depth = depth_map * scale
        absolute_difference_map = np.abs(scaled_depth - self._depth_gt)
        made = np.nanmean(absolute_difference_map)

        if return_scale:
            return made, absolute_difference_map, scale
        else:
            return made, absolute_difference_map