malith26's picture
Deploy to Hugging Face Spaces (2026-09-10 16:04:26 UTC from commit 9042768)
5ed0011
Raw History Blame Contribute Delete
2.56 kB
"""Quantitative evaluation metrics for stereo disparity maps.
Implements RMSE, AbsRel, and Bad-Pixel Rate (D1 metric) conforming to Middlebury benchmark standards.
"""
import numpy as np
from typing import Dict, Tuple, Optional
def calculate_metrics(
gt_disp: np.ndarray,
pred_disp: np.ndarray,
mask: Optional[np.ndarray] = None,
bad_threshold: float = 3.0,
) -> Dict[str, float]:
"""Calculates quantitative benchmark metrics against ground-truth disparity.
Conforms to the final Colab implementation (Cell 22):
- RMSE (Root Mean Squared Error): sqrt(mean((gt - pred)**2))
- AbsRel (Absolute Relative Error): mean(|gt - pred| / gt)
- Bad-pixel rate (D1 error rate): percentage of valid pixels with absolute error > bad_threshold
Args:
gt_disp: Ground-truth disparity map (2D float array).
pred_disp: Predicted disparity map (2D float array).
mask: Optional boolean mask of valid pixels. If None, valid is (gt_disp > 0) & (gt_disp < inf).
bad_threshold: Pixel error threshold for bad pixels (default: 3.0 px for Middlebury).
Returns:
dict containing 'rmse', 'absrel', 'bad_pixel_rate', and 'valid_pixels'.
"""
if mask is None:
mask = (gt_disp > 0) & (gt_disp < np.inf)
# Ensure prediction is resized if dimensions differ slightly (e.g. from downsampling or cropping)
if gt_disp.shape != pred_disp.shape:
import cv2
pred_disp = cv2.resize(pred_disp, (gt_disp.shape[1], gt_disp.shape[0]), interpolation=cv2.INTER_LINEAR)
gt_masked = gt_disp[mask]
pred_masked = pred_disp[mask]
if gt_masked.size == 0:
return {
'rmse': float('nan'),
'absrel': float('nan'),
'bad_pixel_rate': float('nan'),
'valid_pixels': 0
}
# Absolute difference
abs_diff = np.abs(gt_masked - pred_masked)
# RMSE (Root Mean Squared Error)
rmse = float(np.sqrt(np.mean((gt_masked - pred_masked) ** 2)))
# Absolute Relative Error (AbsRel)
absrel = float(np.mean(abs_diff / gt_masked))
# Bad-pixel rate (percentage of pixels with error > bad_threshold)
bad_pixels = np.sum(abs_diff > bad_threshold)
bad_pixel_rate = float((bad_pixels / gt_masked.size) * 100.0)
# Mean Absolute Error / End-Point Error (EPE)
epe = float(np.mean(abs_diff))
return {
'rmse': rmse,
'absrel': absrel,
'bad_pixel_rate': bad_pixel_rate,
'epe': epe,
'valid_pixels': int(gt_masked.size)
}