LiFR-Seg: Anytime High-Frame-Rate Segmentation via Event-Guided Propagation

Xiaoshan Wu, Xiaoyang Lyu, Yifei Yu, Bo Wang, Zhongrui Wang, Xiaojuan Qi
The University of Hong Kong · Southern University of Science and Technology

ICLR 2026

PaperPDFBibTeX

TL;DREvent-guided, uncertainty-aware feature propagation updates semantic predictions at arbitrary query times without a new RGB frame.

Original LiFR-Seg teaser: semantic predictions at interframe query times.
Original LiFR-Seg teaser: semantic predictions at interframe query times. Original PDF ↗

Overview

LiFR-Seg predicts interframe semantic maps from one past RGB frame and observed events. It transports semantic features using event-derived motion and learned confidence, then uses temporal memory to refine them. No future RGB frame is required at inference.

Method

Move the features, not just the pixels.

01

Anchor

Extract semantic features from the RGB keyframe.

02

Propagate

Use event motion and confidence to transport features.

03

Refine

Retrieve temporal context before decoding the semantic map.

Uncertainty-guided splatting and temporal memory. Original Figure 2; future observations shown for supervision are not inference inputs.
Uncertainty-guided splatting and temporal memory. Original Figure 2; future observations shown for supervision are not inference inputs.Click to enlarge ↗

Results

Which representation should move?

ApproachmIoU
Image interpolation70.38
Image warping72.37
Segmentation warping71.63
Feature warping · LiFR-Seg73.82

DSEC, 50 ms query offset. Table 4, mIoU (%). Higher is better. The paper separately reports a target-RGB reference of 73.91% versus 73.82% for LiFR-Seg.

Read the experimental protocol ↗
Interframe segmentation examples. Original Figure 4; target-time ground truth is a reference, not a model input.
Interframe segmentation examples. Original Figure 4; target-time ground truth is a reference, not a model input.Click to enlarge ↗

BibTeX

@inproceedings{wu2026lifrseg,
  title = {LiFR-Seg: Anytime High-Frame-Rate Segmentation via Event-Guided Propagation},
  author = {Xiaoshan Wu and Xiaoyang Lyu and Yifei Yu and Bo Wang and Zhongrui Wang and Xiaojuan Qi},
  booktitle = {International Conference on Learning Representations},
  year = {2026},
  url = {https://arxiv.org/abs/2603.21115}
}