FastGeSCF.

IROS 2026

Towards Practical Scene Change Detection

A Fast, Unaligned Video Framework via Spatiotemporal Alignment

Yi-Heng Yeh · Chuan-Yuan Huang · Kuan-Wen Chen · Li-Yu Lu

Code Paper Video

Abstract

Scene change detection that starts with the real world.

Scene change detection in robot deployments is difficult because repeated traversals are rarely captured at the same time, speed, or viewpoint. FastGeSCF detects meaningful scene changes directly from unaligned video traversals by combining temporal alignment, spatial registration, and selective segmentation.

The result is a practical two-stage pipeline that preserves competitive accuracy while substantially reducing per-pair inference time.

Full pipeline

From unaligned video traversals to change masks.

Complete FastGeSCF pipeline: Stage 1 temporally aligns query and reference videos with VPR descriptors and FastDTW; Stage 2 detects scene changes in the resulting image pairs.
Stage 1 — Temporal Alignment. VPR descriptors and FastDTW identify corresponding frames across unaligned traversals. Stage 2 — Scene Change Detection. FastGeSCF predicts change masks for the aligned image pairs.

FastGeSCF

Spatial alignment, pseudo-mask generation, and selective SAM refinement.

FastGeSCF architecture: query and reference images are spatially aligned, used for cross-attention pseudo-mask generation, and refined into the final scene-change mask.
The FastGeSCF model spatially registers the query and reference images, generates appearance and disappearance pseudo-masks with cross-attention, then refines them through grid-point filtering, SAM prompting, and geometric intersection.

Robust correspondence. Temporal alignment handles uneven sampling and route speed.

Geometrically aware. Spatial registration reduces viewpoint-induced false positives.

Efficient masks. Selective prompting avoids expensive segmentation where it is unnecessary.

Results

Fast enough for practical video-based SCD.

0.658 Average F1
0.888s Average time per pair
7.9× Faster than GeSCF
MethodAvg. F1 ↑Time / pair ↓
GeSCF0.6436.982 s
ZSSCD0.43010.128 s
FastGeSCF0.6580.888 s

Citation

If FastGeSCF helps your research, please cite it.

@inproceedings{yeh2026fastgescf,
  title     = {Towards Practical Scene Change Detection: A Fast, Unaligned Video Framework via Spatiotemporal Alignment},
  author    = {Yi-Heng Yeh and Chuan-Yuan Huang and Kuan-Wen Chen and Li-Yu Lu},
  booktitle = {IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)},
  year      = {2026}
}