基于改进超像素采样的立体匹配网络

Stereo Matching Network Based on Enhanced Superpixel Sampling

针对立体匹配中细节丢失、有遮挡,以及无纹理区域匹配精度低的问题,提出了一种基于改进超像素采样的立体匹配方法。首先,利用改进的超像素采样方法对用于立体匹配的高分辨率输入图像进行下采样,随后,将下采样后的图像对输入到立体匹配网络中,利用权值共享的卷积网络进行特征提取,使用3D卷积获取特征融合后的Cost Volume并生成视差图,再将输出的视差图进行上采样还原为最终的视差图。针对超像素采样过程中容易丢失细节从而影响后续立体匹配精度的问题,引入特征金字塔注意力模块(Feature Pyramid Attention,FPA)和改进的残差结构。根据上述两个方面的创新,提出了基于超像素采样的立体匹配网络FPSMnet(Feature Pyramid Stereo Matching Network),并选取、划分图像数据集BSDS500和NYUv2作为超像素采样的训练、验证和测试的数据集。立体匹配实验结果表明,与基准方法相比,所提算法在SceneFlow和HR-VS数据集上的平均像素误差分别下降了0.25和0.52,在不影响运行时间的前提下提高了匹配精度。

Aiming at the accuracy challenges in stereo matching related to details,occlusion,and textureless regions,a stereo matching method based on improved superpixel sampling is proposed.Initially,an enhanced superpixel sampling method is employed to downsample the high-resolution input images used for stereo matching.Subsequently,the downsampled image pairs are input into the stereo matching network,where a convolutional network with shared weights is utilized for feature extraction.Using 3D convolution,a feature-fused Cost Volume is generated,leading to the creation of a disparity map.The outputted disparity map is then upsampled to reconstruct the final disparity map.To tackle the issue of potential detail loss during the superpixel sampling process,two innovations are introduced:the feature pyramid attention module(FPA)and an improved residual structure.Based on these two innovations,a stereo matching network named FPSMnet(feature pyramid stereo matching network)is proposed.This paper selects and partitions the image datasets BSDS500 and NYUv2 for training,validation,and testing of superpixel sampling.Experimental results in stereo matching demonstrate that,compared to the baseline method,the proposed algorithm achieves a reduction of 0.25 and 0.52 in average pixel errors on the SceneFlow and HR-VS datasets,respectively.These improvements are achieved without compromising runtime efficiency.