LiFR v2: Completion-Augmented Event Propagation for High-Rate Dense Prediction

Tao Wan*, Xiaoshan Wu*†, Yifei Yu, Bo Wang, Xiaoyang Lyu, Muxin Liu, Aoxuan Pan, Zhongrui Wang, Xiaojuan Qi
Southern University of Science and Technology · The University of Hong Kong
* Equal contribution · † Project lead

Under review · IEEE TPAMI

PaperPDFBibTeXRepository

TL;DRPropagation reuses observed content, event-guided completion recovers missing source support, and memory retains inferred content for later queries.

Original project teaser from the official LiFR v2 repository.
Original project teaser from the official LiFR v2 repository. Original PDF ↗

Overview

LiFR v2 supports causal anytime and streaming segmentation, depth and joint prediction. An Event-Guided Completion Module addresses regions missing from the RGB keyframe; a History Retrieval Module reuses completed features. SHF-Emerge tests rapid emergence and disocclusion.

Method

Beyond propagation alone.

01

Propagation

Transport reliable RGB-derived features with scene motion.

02

Completion

Recover event-supported content absent from the keyframe.

03

Retention

Preserve completed representations for subsequent queries.

Streaming architecture, original Figure 2: propagate → retrieve history → complete → decode. Completed features update memory.
Streaming architecture, original Figure 2: propagate → retrieve history → complete → decode. Completed features update memory.Click to enlarge ↗

Results

Recovering newly visible content.

TaskBaselineBaseline scoreLiFR v2
Segmentation ↑LiFR-Seg54.2856.13
Depth ↓LiFR propagation1.5641.118

SHF-Emerge, single-task evaluation. Tables I–II. Segmentation is mIoU (%); depth is 1–10 m RMSE (m). These are different task-specific models.

Read the experimental protocol ↗
Original Figure 8: segmentation, depth and multi-task comparisons. Target-time RGB and HFR outputs are visualization references, not LiFR v2 inputs.
Original Figure 8: segmentation, depth and multi-task comparisons. Target-time RGB and HFR outputs are visualization references, not LiFR v2 inputs.Click to enlarge ↗

BibTeX

@misc{wan2026lifrv2,
  title = {LiFR v2: Completion-Augmented Event Propagation for High-Rate Dense Prediction},
  author = {Tao Wan and Xiaoshan Wu and Yifei Yu and Bo Wang and Xiaoyang Lyu and Muxin Liu and Aoxuan Pan and Zhongrui Wang and Xiaojuan Qi},
  year = {2026},
  eprint = {2609.25803},
  archivePrefix = {arXiv},
  primaryClass = {cs.CV},
  url = {https://arxiv.org/abs/2609.25803}
}

The public repository currently hosts paper materials; its README says code and pretrained models are being prepared.