Models Matter, So Does Training: An Empirical Study of CNNs for Optical Flow Estimation

Area of Science:

• Computer Vision
• Deep Learning

Background:

• Optical flow estimation is crucial for understanding motion in videos.
• Convolutional Neural Networks (CNNs) have shown promise but require efficient models and effective training.

Purpose of the Study:

• To develop a compact and effective CNN model for optical flow estimation.
• To analyze and improve training procedures for enhanced performance.

Main Methods:

• Designed PWC-Net using pyramidal processing, warping, and cost volume.
• Experimentally analyzed performance gains by retraining FlowNetC with new protocols.
• Further refined training procedures to boost accuracy.

Main Results:

• PWC-Net is significantly smaller, faster, and more accurate than FlowNet2.
• Retrained FlowNetC showed substantial accuracy improvements.
• Enhanced training increased PWC-Net accuracy by up to 20% on benchmark datasets.

Conclusions:

• PWC-Net offers a highly efficient and accurate solution for optical flow.
• Optimized training protocols are critical for maximizing CNN performance in computer vision.
• The study provides valuable insights and resources for the research community.