LangGas

01 — Overview

Overview

Gas plumes are semi-transparent, deform over time, and are difficult to label at scale. LangGas addresses both the data shortage and the detection problem through SimGas, a synthetic video dataset with varied scenes, distractors, leak locations, and pixel-level ground truth.

The accompanying method uses language to distinguish plume-like motion from foreground objects, enabling segmentation without task-specific model training.

02 — Contributions

Key Contributions

• 01

  Introduces SimGas, a synthetic gas-leak dataset with precise segmentation masks and diverse scene conditions.

• 02

  Combines enhanced background subtraction with zero-shot object detection and language-based filtering.

• 03

  Uses promptable segmentation and temporal filtering to turn retained detections into stable plume masks.

03 — Method

Method

The pipeline first enhances frame differences produced by background subtraction. A text prompt guides zero-shot object detection, non-maximum suppression and temporal logic remove implausible regions, and SAM 2 segments the remaining candidate plume regions.

04 — Evaluation

Results

On SimGas, the full pipeline reaches an overall IoU of 69%, outperforming baselines based only on background subtraction or zero-shot detection and segmentation. The authors also report qualitative transfer to the real-world GasVid dataset.

05 — Reference

Citation

@InProceedings{Guo_2025_CVPR,
  author    = {Guo, Wenqi and Du, Yiyang and Du, Shan},
  title     = {LangGas: Introducing Language in Selective Zero-Shot Background Subtraction for Semi-Transparent Gas Leak Detection with a New Dataset},
  booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops},
  month     = {June},
  year      = {2025},
  pages     = {4529--4539}
}