Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

SRFlow: A Dataset and Regularization Model for High-Resolution Facial Optical Flow via Splatting Rasterization

JiaLin Zhang, Dong Li

TL;DR

SRFlow introduces a high-resolution facial optical flow dataset built from a dynamic 3D Gaussian avatar and a Flow Rasterizer, enabling dense, structurally consistent supervision for motion. SRFlowNet adopts a backbone selected for subtle facial motion and integrates four regularization losses (TVR, FDR, MIGAR, IGVAR) to suppress noise while preserving fine-grained movement, achieving improved EPE and micro-expression recognition metrics. Training on SRFlow yields substantial gains across optical flow models and ups performance on micro-expression benchmarks, with up to 42% reduction in EPE and up to 48% improvement in macro F1 on a composite micro-expression dataset. Together, the dataset and regularized model demonstrate that high-resolution facial flow supervision can substantially enhance both low-level motion estimation and downstream expressive understanding in real-world scenarios.

Abstract

Facial optical flow supports a wide range of tasks in facial motion analysis. However, the lack of high-resolution facial optical flow datasets has hindered progress in this area. In this paper, we introduce Splatting Rasterization Flow (SRFlow), a high-resolution facial optical flow dataset, and Splatting Rasterization Guided FlowNet (SRFlowNet), a facial optical flow model with tailored regularization losses. These losses constrain flow predictions using masks and gradients computed via difference or Sobel operator. This effectively suppresses high-frequency noise and large-scale errors in texture-less or repetitive-pattern regions, enabling SRFlowNet to be the first model explicitly capable of capturing high-resolution skin motion guided by Gaussian splatting rasterization. Experiments show that training with the SRFlow dataset improves facial optical flow estimation across various optical flow models, reducing end-point error (EPE) by up to 42% (from 0.5081 to 0.2953). Furthermore, when coupled with the SRFlow dataset, SRFlowNet achieves up to a 48% improvement in F1-score (from 0.4733 to 0.6947) on a composite of three micro-expression datasets. These results demonstrate the value of advancing both facial optical flow estimation and micro-expression recognition.

SRFlow: A Dataset and Regularization Model for High-Resolution Facial Optical Flow via Splatting Rasterization

TL;DR

SRFlow introduces a high-resolution facial optical flow dataset built from a dynamic 3D Gaussian avatar and a Flow Rasterizer, enabling dense, structurally consistent supervision for motion. SRFlowNet adopts a backbone selected for subtle facial motion and integrates four regularization losses (TVR, FDR, MIGAR, IGVAR) to suppress noise while preserving fine-grained movement, achieving improved EPE and micro-expression recognition metrics. Training on SRFlow yields substantial gains across optical flow models and ups performance on micro-expression benchmarks, with up to 42% reduction in EPE and up to 48% improvement in macro F1 on a composite micro-expression dataset. Together, the dataset and regularized model demonstrate that high-resolution facial flow supervision can substantially enhance both low-level motion estimation and downstream expressive understanding in real-world scenarios.

Abstract

Facial optical flow supports a wide range of tasks in facial motion analysis. However, the lack of high-resolution facial optical flow datasets has hindered progress in this area. In this paper, we introduce Splatting Rasterization Flow (SRFlow), a high-resolution facial optical flow dataset, and Splatting Rasterization Guided FlowNet (SRFlowNet), a facial optical flow model with tailored regularization losses. These losses constrain flow predictions using masks and gradients computed via difference or Sobel operator. This effectively suppresses high-frequency noise and large-scale errors in texture-less or repetitive-pattern regions, enabling SRFlowNet to be the first model explicitly capable of capturing high-resolution skin motion guided by Gaussian splatting rasterization. Experiments show that training with the SRFlow dataset improves facial optical flow estimation across various optical flow models, reducing end-point error (EPE) by up to 42% (from 0.5081 to 0.2953). Furthermore, when coupled with the SRFlow dataset, SRFlowNet achieves up to a 48% improvement in F1-score (from 0.4733 to 0.6947) on a composite of three micro-expression datasets. These results demonstrate the value of advancing both facial optical flow estimation and micro-expression recognition.
Paper Structure (23 sections, 19 equations, 3 figures, 10 tables)

This paper contains 23 sections, 19 equations, 3 figures, 10 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Optical Flow Estimation
  4. Optical Flow Datasets
  5. Facial Optical Flow
  6. SRFlow Dataset
  7. 3D Gaussian Avatar Reconstruction
  8. 3D Gaussian Dataset Generation
  9. SRFlowNet Model
  10. Model Backbone
  11. Regularization Loss for High-Resolution Facial Optical Flow
  12. Total Variation Regularization Loss (TVR)
  13. Flow Difference Regularization Loss (FDR)
  14. Mean Image Gradient Activation Regularization Loss (MIGAR)
  15. Image Gradient Variance Activation Regularization Loss (IGVAR)
  16. ...and 8 more sections

Figures (3)

  • Figure 1: Illustration of Optical Flow Datasets Featuring Facial Motion
  • Figure 2: Overall Pipeline for Dataset Generation and Network Training
  • Figure 3: Qualitative Visualization of Optical Flow Estimations Across All Evaluated Networks.