From 6e7357b2507b07a40bb7a102e552b0e97e13cb14 Mon Sep 17 00:00:00 2001
From: Vladyslav Usenko <vlad.usenko@tum.de>
Date: Thu, 11 Jul 2019 18:49:13 +0200
Subject: [PATCH] updated gamma correction script

---
 scripts/check_gamma_correction.py | 63 ++++++++++++++++++++++++++++---
 1 file changed, 57 insertions(+), 6 deletions(-)

diff --git a/scripts/check_gamma_correction.py b/scripts/check_gamma_correction.py
index 161f4bb..192dff1 100755
--- a/scripts/check_gamma_correction.py
+++ b/scripts/check_gamma_correction.py
@@ -3,9 +3,8 @@
 import sys
 import math
 import os
-import webp
+import cv2
 
-import pandas as pd
 import numpy as np
 from matplotlib import pyplot as plt
 
@@ -14,20 +13,72 @@ dataset_path = sys.argv[1]
 print(dataset_path)
 
 timestamps = np.loadtxt(dataset_path + '/mav0/cam0/data.csv', usecols=[0], delimiter=',', dtype=np.int64)
-exposures = np.loadtxt(dataset_path + '/mav0/cam0/exposure.csv', usecols=[1], delimiter=',', dtype=np.int64)
+exposures = np.loadtxt(dataset_path + '/mav0/cam0/exposure.csv', usecols=[1], delimiter=',', dtype=np.int64).astype(np.float64) * 1e-9
 pixel_avgs = list()
 
+if timestamps.shape[0] != exposures.shape[0]: print("timestamps and exposures do not match")
+
+imgs = []
 
 # check image data.
-img_extensions = ['.png', '.jpg', '.webp']
 for timestamp in timestamps:
     path = dataset_path + '/mav0/cam0/data/' + str(timestamp)
-    img = webp.imread(dataset_path + '/mav0/cam0/data/' + str(timestamp) + '.webp')
+    img = cv2.imread(dataset_path + '/mav0/cam0/data/' + str(timestamp) + '.webp', cv2.IMREAD_GRAYSCALE)[:,:,0]
+    imgs.append(img)
     pixel_avgs.append(np.mean(img))
 
-plt.plot(exposures, pixel_avgs)
+imgs = np.array(imgs)
+print(imgs.shape)
+print(imgs.dtype)
+
+inv_resp = np.arange(256, dtype=np.float64)
+inv_resp[250:] = -1.0 # Use negative numbers to detect oversaturation 
+
+irradiance = imgs[0] / exposures[0]
+
+def opt_irradiance():
+    corrected_imgs = inv_resp[imgs] * exposures[:, np.newaxis, np.newaxis]
+    times = np.ones_like(corrected_imgs) * (exposures**2)[:, np.newaxis, np.newaxis]
+    times[corrected_imgs < 0] == 0
+    corrected_imgs[corrected_imgs < 0] == 0
+
+    irr = np.sum(corrected_imgs, axis=0) / np.sum(times, axis=0)
+    return irr
+
+def opt_inv_resp():
+    generated_imgs = irradiance[np.newaxis, :, :] * exposures[:, np.newaxis, np.newaxis]
+    
+    num_pixels_by_intensity = np.bincount(imgs.flat)
+    sum_by_intensity = np.bincount(imgs.flat, generated_imgs.flat)
+    
+    new_inv_resp = inv_resp
+
+    idx = np.nonzero(num_pixels_by_intensity > 0)
+    new_inv_resp[idx] = sum_by_intensity[idx] / num_pixels_by_intensity[idx]
+    new_inv_resp[250:] = -1.0
+    return new_inv_resp
+
+def print_error():
+    generated_imgs = irradiance[np.newaxis, :, :] * exposures[:, np.newaxis, np.newaxis]
+    generated_imgs -= inv_resp[imgs]
+    generated_imgs[imgs == 255] = 0
+    print(np.sum(generated_imgs**2))
+
+print_error()
+for iter in range(3):
+    irradiance = opt_irradiance()
+    print_error()
+    inv_resp = opt_inv_resp()
+    print_error()
+
+
+plt.figure()
+plt.plot(inv_resp)
 plt.ylabel('Img Mean')
 plt.xlabel('Exposure')
+
+plt.figure()
+plt.imshow(irradiance)
 plt.show()