基于层级特征融合的单目深度估计算法

MonoDepth2的提出使自监督单目深度估计取得了 重大的进展,但该网络在大的无语义区域和边界处预测效果并不理想, 主要原因是基础的U-Net框架没有充分利用多尺度特征信息,导致来自于大梯度区域的深 度估计较差。针对此问题,本文提出 了一个改进的DepthNet,层级特征融合网络(hierarchical integration net,HINet)。优 化了U-Net网络结构,使编码器端在每一层 都能产生不同尺度的特征信息,从而让解码器端在每一层都能够充分融合多尺度特征。由于 不同尺度的特征信息对于特定的解 码器层都有不同程度的贡献,本文提出的层级特征融合算法还增加了通道注意力模块,提升 重要特征尺度的权重。当采用立体 图像对进行训练时,本文对数据进行了预处理,并增加了立体对的深度暗示损失函数。在KI TTI 数据集上的实验结果表明,所 有指标均获得了不同程度的提升,其中绝对相对误差减少了0.09,平 方相对误差减少了0.093。

Monocular depth estimation algorithm based on hierarchical integration

The proposal of MonoDepth2 has made significant progress in self-supe rvised monocular depth estimation,but the prediction effect of the network in large non semantic regions and boundarie s is not ideal.The main reason is that the basic U-Net framework does not make full use of multi-scale feature informatio n,resulting in poor depth estimation from large gradient regions.To address this problem,this paper proposed an improved DepthNet,a hierarchical integration net (HINet).The U-Net network structure is optimized so that the encoder side can generate feature information of different scales at each layer,thus allowing the decoder side to fully fuse multi-scale features at each layer.Since the feature information of different scales contributes to a specific decoder layer to diffe rent degrees,the hierarchical integration (HINet) algorithm proposed in this paper also adds a channel attention module to enhance the weight of important feature scales.When stereo pairs are used for training,this paper preprocesses the dat a and adds a depth-implying loss function for stereo pairs.The experimental results on the KITTI dataset show that all indicators are improved to varying degrees,in which the absolute relative error is reduced by 0.09 and the squared relative error is reduced by 0.093.