FGSTP

01 — Overview

Overview

Gas leaks have weak texture, translucent boundaries, and highly variable motion. FGSTP combines temporal correspondence with local spatial detail so the model can follow plume movement without losing fine boundaries.

The work also introduces GasVid, a manually annotated video dataset created for evaluating gas-leak segmentation.

02 — Contributions

Key Contributions

• 01

  Builds a correlation volume across consecutive frames to expose motion cues from subtle plume movement.

• 02

  Refines spatial detail and boundaries through a dedicated fine-grained decoder.

• 03

  Introduces a manually labeled gas-leak video dataset for training and evaluation.

03 — Method

Method

FGSTP extracts features from adjacent frames, computes their correlation volume, and combines the resulting temporal evidence with spatial features. A boundary-aware decoder progressively refines the predicted gas mask.

04 — Evaluation

Results

Experiments on GasVid show that combining motion correlation and spatial refinement improves segmentation accuracy, particularly around faint and irregular plume boundaries.

05 — Reference

Citation

@InProceedings{Zhao_2025_FGSTP,
  author    = {Zhao, Xinlong and Du, Shan},
  title     = {Fine-Grained Spatial-Temporal Perception for Gas Leak Segmentation},
  booktitle = {2025 IEEE International Conference on Image Processing (ICIP)},
  year      = {2025},
  pages     = {869--874},
  doi       = {10.1109/ICIP55913.2025.11084304}
}