PFGNet: A Fully Convolutional Frequency-Guided Peripheral Gating Network for Efficient Spatiotemporal Predictive Learning

PFGNet is a fully convolutional framework for efficient spatiotemporal predictive learning (STPL), presented at CVPR 2026. It aims to forecast future frames from past observations by dynamically modulating receptive fields through pixel-wise frequency-guided gating.

Inspired by biological center-surround organization, the core Peripheral Frequency Gating (PFG) block extracts localized spectral cues to adaptively fuse multi-scale large-kernel peripheral responses with learnable center suppression, forming spatially adaptive band-pass filters.

Resources:

• Paper: PFGNet: A Fully Convolutional Frequency-Guided Peripheral Gating Network for Efficient Spatiotemporal Predictive Learning
• Code: Official GitHub Repository
• Project Page: kaimaoge.github.io

Available checkpoints

This repository provides dataset-specific trained checkpoints of PFGNet on multiple benchmarks:

Dataset	Checkpoint
Moving MNIST	pfg_mmnist.ckpt
Moving Fashion MNIST	pfg_mfmnist.ckpt
TaxiBJ	pfg_taxibj.ckpt
KTH (10→20)	pfg_kth20.ckpt
KTH (10→40)	pfg_kth40.ckpt
Human3.6M	pfg_human.ckpt

Usage

PFGNet directly inherits the codebase and dependencies of OpenSTL. Please refer to the official repository for detailed environment setup and data preparation instructions.

Training (Moving MNIST example)

From the repository root, run:

python tools/train.py -d mmnist -c configs/mmnist/PFG.py --ex_name mmnist_pfg --test

Testing (Moving MNIST example)

From the repository root, run:

python tools/test.py -d mmnist -c configs/mmnist/PFG.py --ex_name mmnist_pfg --test

Citation

If you find this work helpful, please consider citing:

@misc{cai2026pfgnetfullyconvolutionalfrequencyguided,
      title={PFGNet: A Fully Convolutional Frequency-Guided Peripheral Gating Network for Efficient Spatiotemporal Predictive Learning}, 
      author={Xinyong Cai and Changbin Sun and Yong Wang and Hongyu Yang and Yuankai Wu},
      year={2026},
      eprint={2602.20537},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}