An improved LiteFlowNet3 method for precise optical flow calculation in video

Abstract

The paper proposes an improved LiteFlowNet3 method for precise optical flow calculation in video. It combines probabilistic modeling with adversarial training and improves the cost-volume modulation module of LiteFlowNet3. By introducing a Gaussian Mixture Model (GMM) to model the distribution of optical flow residuals and designing a generative adversarial mechanism to discriminate anomalous regions, the proposed method achieves end-to-end filtering of outliers. Experimental evaluations on MPI-Sintel optical flow datasets demonstrate that the proposed method reduces the average endpoint error (EPE) by 14.3% and enhances the anomaly detection AUC to 0.88, significantly outperforming the baseline LiteFlowNet3. The approach enhances the robustness of optical flow estimation for practical applications such as video analysis