其他分享
首页 > 其他分享> > 《Boosting Monocular Depth Estimation Models to High-Resolution...》论文笔记

《Boosting Monocular Depth Estimation Models to High-Resolution...》论文笔记

作者:互联网

参考代码:BoostingMonocularDepth
论文:Boosting Monocular Depth Estimation Models to High-Resolution via Content-Adaptive Multi-Resolution Merging

1. 概述

导读:这篇文章提出了一种使用现有深度估计模型(MiDas)生成具有更高分辨率/更多细节表现深度图的方法(同时保持高分辨率和深度一致性是现存的挑战)。文章探究了输入的图像分辨率/图像中的结构信息是如何影响最后的深度估计的,也就是输入图像分辨率对模型带来的影响,以及通过图像中的结构信息进行细粒度深度细节优化。文章通过选择合适的策略在保持场景预测连续性和高频信息表达能力上取折中,生成了具有更多细节表现能力的深度估计方法。文章的策略主要归纳为两点:
1)探究分辨率和模型生成深度图的关系,提出一种高低分辨率结果融合机制,提升深度估计的细节表达;
2)在一些细节上使用patch采样的方式对细节信息进行补全,进一步提升细节上的表达,同时也提出一种path和原图的融合策略,用以消除融合问题(path和原图预测出来的深度不一致)。

1.1 对于深度估计模型的观察

输入分辨率与深度估计结果:
在低分辨率(也就是训练使用的分辨率)下,生成的深度预测结果具有较好的深度一致性,但是当把分辨率加大的时候,在出现更多深度细节的同时,深度的一致性预测也被破坏。可以参见下图所示:
在这里插入图片描述

细节分布与深度估计结果:
在大尺寸的深度图预测中,出于CNN网络感受野的原因,在深度细节比CNN网络的感受野相差较远的时候那么就会出现深度预测的不一致性,也就是图像中不同区域的适宜分辨率是各异的。对此文章中通过图像进行采样的方式,为每个采样选择合适的输入分辨率,从而得到更好的深度预测结果。

对与上面的两点文章对其进行分析,将出现这些问题的原因归纳为如下的两点:

对此,文章通过上述中提到的不同分辨率深度结果融合/patch选择优化之后融合,从而得到高分辨率且细节表现丰富的深度预测结果。其效果可以参考下图所示:
在这里插入图片描述

2. 方法设计

2.1 整体pipline

在这里插入图片描述
文章整体的pipeline流程可以划分为如下的几个步骤:

2.2 Double Estimation策略

这里首先是输入两个不同分辨率的图像作为输入(也就是训练分辨率和最大无一致性问题的输入分辨率)得到两个深度预测。那么对于这个最大无问题的分辨率(也就是文章中提到的 R 20 R_{20} R20​)是怎么确定的呢?文章是通过原模型感受野作为边界,按照一定的规则进行选择的,其过程参考:

# utils.py:70
def calculateprocessingres(img, basesize, confidence=0.1, scale_threshold=3, whole_size_threshold=3000):
    # Returns the R_x resolution described in section 5 of the main paper.

    # Parameters:
    #    img :input rgb image
    #    basesize : size the dilation kernel which is equal to receptive field of the network.
    #    confidence: value of x in R_x; allowed percentage of pixels that are not getting any contextual cue.
    #    scale_threshold: maximum allowed upscaling on the input image ; it has been set to 3.
    #    whole_size_threshold: maximum allowed resolution. (R_max from section 6 of the main paper)

    # Returns:
    #    outputsize_scale*speed_scale :The computed R_x resolution
    #    patch_scale: K parameter from section 6 of the paper

    # speed scale parameter is to process every image in a smaller size to accelerate the R_x resolution search
    speed_scale = 32  # 加快R_x的计算
    image_dim = int(min(img.shape[0:2]))

    gray = rgb2gray(img)  # 计算梯度
    grad = np.abs(cv2.Sobel(gray, cv2.CV_64F, 0, 1, ksize=3)) + np.abs(cv2.Sobel(gray, cv2.CV_64F, 1, 0, ksize=3))
    grad = cv2.resize(grad, (image_dim, image_dim), cv2.INTER_AREA)

    # thresholding the gradient map to generate the edge-map as a proxy of the contextual cues
    m = grad.min()
    M = grad.max()
    middle = m + (0.4 * (M - m))
    grad[grad < middle] = 0
    grad[grad >= middle] = 1

    # dilation kernel with size of the receptive field
    kernel = np.ones((int(basesize/speed_scale), int(basesize/speed_scale)), np.float)
    # dilation kernel with size of the a quarter of receptive field used to compute k
    # as described in section 6 of main paper
    kernel2 = np.ones((int(basesize / (4*speed_scale)), int(basesize / (4*speed_scale))), np.float)

    # Output resolution limit set by the whole_size_threshold and scale_threshold.
    threshold = min(whole_size_threshold, scale_threshold * max(img.shape[:2]))

    outputsize_scale = basesize / speed_scale
    for p_size in range(int(basesize/speed_scale), int(threshold/speed_scale), int(basesize / (2*speed_scale))):
        grad_resized = resizewithpool(grad, p_size)
        grad_resized = cv2.resize(grad_resized, (p_size, p_size), cv2.INTER_NEAREST)
        grad_resized[grad_resized >= 0.5] = 1
        grad_resized[grad_resized < 0.5] = 0

        dilated = cv2.dilate(grad_resized, kernel, iterations=1)  # 按照感受野进行膨胀
        meanvalue = (1-dilated).mean()  # 随着尺寸的变大,计算不在感受野中像素占据的比例
        if meanvalue > confidence:  # 图像中存在太多的不在感受野中的像素
            break
        else:
            outputsize_scale = p_size

    grad_region = cv2.dilate(grad_resized, kernel2, iterations=1)
    patch_scale = (1-grad_region).mean()  # 后面patch确定用到的scale

    return int(outputsize_scale*speed_scale), patch_scale

从上面的过程中确定下来最大的分辨率(也就是文章中提到的 R 20 R_{20} R20​),之后就是对这两个分辨率的图像进行深度估计并融合,参考:

# run.py:342
# Generate a double-input depth estimation
def doubleestimate(img, size1, size2, pix2pixsize, net_type):
    # Generate the low resolution estimation
    estimate1 = singleestimate(img, size1, net_type)
    # Resize to the inference size of merge network.
    estimate1 = cv2.resize(estimate1, (pix2pixsize, pix2pixsize), interpolation=cv2.INTER_CUBIC)

    # Generate the high resolution estimation
    estimate2 = singleestimate(img, size2, net_type)
    # Resize to the inference size of merge network.
    estimate2 = cv2.resize(estimate2, (pix2pixsize, pix2pixsize), interpolation=cv2.INTER_CUBIC)

    # Inference on the merge model
    pix2pixmodel.set_input(estimate1, estimate2)
    pix2pixmodel.test()
    visuals = pix2pixmodel.get_current_visuals()
    prediction_mapped = visuals['fake_B']
    prediction_mapped = (prediction_mapped+1)/2
    prediction_mapped = (prediction_mapped - torch.min(prediction_mapped)) / (
                torch.max(prediction_mapped) - torch.min(prediction_mapped))
    prediction_mapped = prediction_mapped.squeeze().cpu().numpy()

    return prediction_mapped

在这里插入图片描述
在上图中(a)是原分辨率预测的结果,(c)是按照文章最大分辨率预测得到的结果,(d)是融合的结果。

2.3 Patch Estimates策略

在完成上述的double estimates之后,文章还通过patch选择的方式进行细节优化,其中对于patch的选择与上述内容中 R 20 R_{20} R20​的确定过程类似,只不过是多了一些patch排除策略。首先计算宽度上的梯度均值和梯度积分图:

# run.py:267
# We use the integral image to speed-up the evaluation of the amount of gradients for each patch.
gf = whole_grad.sum()/len(whole_grad.reshape(-1))
grad_integral_image = cv2.integral(whole_grad)

之后按照策略进行无效patch排除:

# run.py:303
# Compute the amount of gradients present in the patch from the integral image.
cgf = getGF_fromintegral(integral_grad, bbox)/(bbox[2]*bbox[3])

# Check if patching is beneficial by comparing the gradient density of the patch to
# the gradient density of the whole image
if cgf >= gf:
	# patch确定与校验
	....

在得到patch之后按照上述内容中的double estimates进行预测,得到深度结果,之后再与原来的深度预测结果进行融合。若不融合的话会出现下面的深度不统一的问题。
在这里插入图片描述
这里的话文章是采取的比较简单的线性融合策略(要理解下面的代码还需要往上再看几行):

# run.py:237
# Update the whole estimation:
# We use a simple Gaussian mask to blend the merged patch region with the base estimate to ensure seamless
# blending at the boundaries of the patch region.
tobemergedto[h1:h2, w1:w2] = np.multiply(tobemergedto[h1:h2, w1:w2], 1 - mask) + np.multiply(merged, mask)

3. 实验结果

在这里插入图片描述

标签:...,scale,Monocular,Models,分辨率,patch,深度,grad,size
来源: https://blog.csdn.net/m_buddy/article/details/117912071