LiteFlowNet is a lightweight, fast, and accurate convolutional neural network for optical flow estimation, targeting researchers and practitioners in computer vision. It offers a smaller model size and competitive performance compared to other state-of-the-art methods, enabling efficient deployment on resource-constrained environments.

How It Works

LiteFlowNet employs a cascaded architecture with several specialized modules: pyramidal features for multi-scale processing, a cost volume and sub-pixel refinement for accurate flow estimation, a feature warping (f-warp) layer to align features across frames, and feature-driven local convolution (f-lconv) for flow regularization. This design balances accuracy with computational efficiency.

Quick Start & Requirements

• Install: Compile from source using make -j 8 after configuring Makefile.config.
• Prerequisites: Ubuntu 14.04/16.04, CUDA 8.0, cuDNN 5.1, OpenCV 2.4.8/3.1.0. Requires manual setup for cuDNN versions.
• Datasets: FlyingChairs (31GB), Things3D (37GB RGB, 311GB flow), Sintel (5.3GB), KITTI12/15 (2GB each).
• Models: Pre-trained models for Sintel and KITTI benchmarks are provided.
• Docs: Project page: http://mmlab.ie.cuhk.edu.hk/projects/LiteFlowNet/

Highlighted Details

• Outperforms PWC-Net on KITTI benchmarks with ~40% smaller model size.
• Achieves 3.27% EPE on KITTI12 and 9.38% EPE on KITTI15.
• Introduces novel f-warp and f-lconv layers for improved feature alignment and regularization.
• Extended versions LiteFlowNet2 and LiteFlowNet3 offer further accuracy and speed improvements.

Maintenance & Community

The original repository is from 2018. Extended versions (LiteFlowNet2, LiteFlowNet3) are available on the same GitHub organization, indicating ongoing research. No specific community channels (Discord/Slack) are mentioned.

Licensing & Compatibility

Provided for academic research purposes only. Commercial use requires explicit consent. The license is not explicitly stated but implies a restrictive, non-commercial use.

Limitations & Caveats

The original Caffe implementation has specific, older dependency requirements (CUDA 8.0, cuDNN 5.1) that may be challenging to set up on modern systems. PyTorch and TensorFlow reimplementations are available for easier integration.