Abstract:Image processing-based optical flow computation has been an appalling task so far. Nowadays, convolution neural network (CNNs), a deep learning method, is broadly applied for estimating optical flow. One of the frameworks of CNNs is the U-Net architecture. This architecture is comprised of encoder-decoder framework that can be trained endways. However, its encoder component utilizes an identical image categorization technique which is common in other categorization architectures. Moreover, the decoder unit is employed to augment the spatial feature maps to full scale of intake by executing successive deconvolution. This type of architecture yields blurred flow fields owing to unpolished features and low-resolution as it is noteworthy that optical flow is pixel-level stint instead. In this article, it has been strived to find a way out for this problem. For this, two architectures, dilated convolution neural network, and channel attention mechanism are introduced inside FlowNetCorr network to estimate optical flow and training loss. In our framework, dilation convolution is deployed to attain spatial precision, and also elaborate the receptive field without requiring huge computational means, and keeps the spatial resolution of feature map unmodified, while channel attention is founded on squeeze-and-excitation architecture for image categorization, which can accommodatingly readjust channel-wise feature by applying global channel knowledge. Comprehensive experiments are executed on MPI-Sintel (Clean and Final), and KITTI (2012 and 2015) test datasets to assess the efficiency of our framework. The experimental outcome denotes that our framework has gone par in performance for MPI-Sintel (Clean and Final) datasets in terms of minimizing training loss, accuracy, and visual betterment to many unsupervised methods, e.g., USCNN, UnSupFlownet, DSTFlow, etc., and cutting-edge supervised methods, e.g., SpyNet, SpyNet+ft, CaF-Full-41c, etc. However, having degraded performance than FlowNet2, while for KITTI (2012 and 2015) datasets, our framework has also achieved better performances over many methods except from UnFlow+ft. These consequences affirm the significance of dilated convolution, and channel attention strategies as well for estimating optical flow.

Guided Learning for Efficient Optical Flow Estimation

A Dense Optical Flow Registration Algorithm Based on Deep Learning

FastFlowNet: A Lightweight Network for Fast Optical Flow Estimation

Learning Optical Flow Using Deep Dilated Residual Networks

Optical Flow Estimation Using Channel Attention Mechanism and Dilated Convolutional Neural Networks.

Learning Optical Flow via Deformable Convolution and Feature Pyramid Networks

Learning By Analogy: Reliable Supervision From Transformations For Unsupervised Optical Flow Estimation

Optical Flow Estimation with Convolutional Neural Nets

Recurrent Partial Kernel Network for Efficient Optical Flow Estimation

A CNNs-based method for optical flow estimation with prior constraints and stacked U-Nets

OPTICAL FLOW ESTIMATION USING CHANNEL ATTENTION MECHANISM

Fdflownet: Fast Optical Flow Estimation Using A Deep Lightweight Network

RFRFlow: Recurrent feature refinement network for optical flow estimation

Optical Flow Estimation with Foreground Attention Guided Network

Ad-Net: Attention Guided Network For Optical Flow Estimation Using Dilated Convolution

CompactFlowNet: Efficient Real-time Optical Flow Estimation on Mobile Devices

Context-Aware Iteration Policy Network for Efficient Optical Flow Estimation

MRDFlow: Unsupervised Optical Flow Estimation Network With Multi-Scale Recurrent Decoder

SKFlow: Optical Flow Estimation Using Selective Kernel Networks

Self-Attention-Based Multiscale Feature Learning Optical Flow with Occlusion Feature Map Prediction

FlowNet: Learning Optical Flow with Convolutional Networks