如何使用纹理改善 StereoSGBM 结果

Question

所以我正在为一个高中项目的 6Dof 做立体 360。我的视差结果不错，但我想知道是否有办法让它们变得更好，特别是它们如何处理纹理。随着点越来越远，视差图应该逐渐消失，但是，因为 StereoSGBM 不能很好地处理纹理，所以远处的点不合理地接近。天空也应该是黑色的，但它很亮。

视差图：

我正在使用 StereoSGBM 获取 2 个 Ricoh Theta SC 相机的视差图。我已经尝试调整视差设置并使用输入图像的亮度和对比度进行播放。我还尝试过更改 StereoSGBM 模式（HH、SGBM、SGBM_3WAY）、翻转输入图像以及使用 StereoBM 而不是 SGBM。我没有尝试校准相机（除了移动图像使相机指向完全相同的方向），因为我认为如果相机校准是一个问题，我会得到更糟糕的结果。除了 StereoSGBM 之外，还有其他一些我可以使用的视差函数可以让我获得更好的结果。我应该尝试将机器学习与 Google Cloud 结合使用来创建更好的视差模型吗？我可以给我们纽约大学深度数据集（https://cs.nyu.edu/~silberman/datasets/nyu_depth_v2.html) 来训练模型。有没有人对如何改善我的结果有任何想法？

import numpy as np
import cv2 as cv
from sklearn.preprocessing import normalize
from PIL import Image, ImageEnhance, ImageOps


def func_disparity(window_size, minDisparity2, numDisparities2, blockSize2, 
    disp12MaxDiff2, uniquenessRatio2, speckleWindowSize2, speckleRange2, 
    preFilterCap2, brightness,
    contrast, event=None):

    imgR = Image.open(FILE_NAME)
    imgL = Image.open(FILEN_NAME)
    print(imgL.size)

    imgL = ImageOps.expand(imgL, border=50)
    imgR = ImageOps.expand(imgR, border=50)    
    
    contrastL = ImageEnhance.Contrast(imgL)
    contrastR = ImageEnhance.Contrast(imgR)
    imgL = contrastL.enhance(contrast)
    imgR = contrastR.enhance(contrast)

    brightnessL = ImageEnhance.Brightness(imgL)
    brightnessR = ImageEnhance.Brightness(imgR)
    imgL = brightnessL.enhance(brightness)
    imgR = brightnessR.enhance(brightness)

    imgL = imgL.convert('L')
    imgL = np.array(imgL)
    imgR = imgR.convert('L')
    imgR = np.array(imgR)
    
    
    #window_size = 15                     wsize default 3; 5; 7 for SGBM reduced size image; 15 for SGBM full size image (1300px and above); 5 Works nicely
 
    left_matcher = cv.StereoSGBM_create(
        minDisparity=minDisparity2,
        numDisparities=numDisparities2,             # max_disp has to be 
        dividable by 16 f. E. HH 192, 256
        blockSize= blockSize2,
        P1=8 * 3 * window_size ** 2,    # wsize default 3; 5; 7 for SGBM 
        reduced size image; 15 for SGBM full size image (1300px and above); 5 
        Works nicely
        P2=32 * 3 * window_size ** 2,
        disp12MaxDiff=disp12MaxDiff2,
        uniquenessRatio=uniquenessRatio2,
        speckleWindowSize=speckleWindowSize2,
        speckleRange=speckleRange2,
        preFilterCap= preFilterCap2,
        mode=cv.STEREO_SGBM_MODE_SGBM_3WAY
        )

    right_matcher = cv.ximgproc.createRightMatcher(left_matcher)

    # FILTER Parameters
    lmbda = 80000
    sigma = 1.2
    visual_multiplier = 1.0
 
    wls_filter = 
    cv.ximgproc.createDisparityWLSFilter(matcher_left=left_matcher)
    wls_filter.setLambda(lmbda)
    wls_filter.setSigmaColor(sigma)

    print('computing disparity...')
    displ = left_matcher.compute(imgL, imgR)  # .astype(np.float32)/16
    dispr = right_matcher.compute(imgR, imgL)  # .astype(np.float32)/16
    displ = np.int16(displ)
    dispr = np.int16(dispr)
    filteredImg = wls_filter.filter(displ, imgL, None, dispr)  # important to 
    put "imgL" here!!!

    filteredImg = cv.normalize(src=filteredImg, dst=filteredImg, beta=0, 
    alpha=255, norm_type=cv.NORM_MINMAX);
    filteredImg = np.uint8(filteredImg)

    height, width = filteredImg.shape
    filteredImg = np.delete(filteredImg, np.s_[0:50], axis=0)
    filteredImg = np.delete(filteredImg, np.s_[height-100:height-50], axis=0)
    filteredImg = np.delete(filteredImg, np.s_[0:50], axis=1)
    filteredImg = np.delete(filteredImg, np.s_[width-100:width-50], axis=1)
    print(filteredImg.shape)
    
    return(filteredImg)
    #print(filteredImg)
    #file = Image.fromarray(filteredImg)
    #file.save("disparitymap.jpg")
    print("Done")

window_size = 3
minDisparity2 = 15
numDisparities2=16      #160max_disp has to be dividable by 16 f. E. HH 192, 
blockSize2=20   #Maybe 20 is optimal
disp12MaxDiff2=1
uniquenessRatio2=15
speckleWindowSize2=0
speckleRange2=2
preFilterCap2=63
brightness=1
contrast=1

disparity = func_disparity(window_size, minDisparity2, numDisparities2, blockSize2, disp12MaxDiff2, uniquenessRatio2, speckleWindowSize2, speckleRange2, preFilterCap2, brightness, contrast, event=None)

file = Image.fromarray(disparity)
file.show()