--- dataset_info: features: - name: filename dtype: string - name: video dtype: string - name: frames dtype: image - name: gt dtype: image splits: - name: train num_bytes: 667056578.338 num_examples: 11758 download_size: 747765689 dataset_size: 667056578.338 configs: - config_name: default data_files: - split: train path: data/train-* --- This repository contains the data presented in [LangGas: Introducing Language in Selective Zero-Shot Background Subtraction for Semi-Transparent Gas Leak Detection with a New Dataset](https://huggingface.co/papers/2503.02910). # LangGas: Introducing Language in Selective Zero-Shot Background Subtraction for Semi-Transparent Gas Leak Detection with a New Dataset Wenqi Guo, Yiyang Du, Shan Du University of British Columbia Group of Methane Emission Observation and Watch (MEOW), Weasoft Software [![arXiv](https://img.shields.io/badge/arXiv-2503.02910-b31b1b.svg)](https://arxiv.org/abs/2503.02910) [![PWC](https://img.shields.io/endpoint.svg?url=https://paperswithcode.com/badge/langgas-introducing-language-in-selective/segmentation-on-simgas)](https://paperswithcode.com/sota/segmentation-on-simgas?p=langgas-introducing-language-in-selective) [Latina](latin_readme.md) [繁体中文](traditional_chinrse_readme.md) [Mewspeak](mewspeak_readme.md) ## Abstract Gas leakage poses a significant hazard that requires prevention. Traditionally, human inspection has been used for detection, a slow and labour-intensive process. Recent research has applied machine learning techniques to this problem, yet there remains a shortage of high-quality, publicly available datasets. This paper introduces a synthetic dataset, SimGas, featuring diverse backgrounds, interfering foreground objects, diverse leak locations, and precise segmentation ground truth. We propose a zero-shot method that combines background subtraction, zero-shot object detection, filtering, and segmentation to leverage this dataset. Experimental results indicate that our approach significantly outperforms baseline methods based solely on background subtraction and zero-shot object detection with segmentation, reaching an IoU of 69% overall. We also present an analysis of various prompt configurations and threshold settings to provide deeper insights into the performance of our method. ## Method Overview ![meow2](https://github.com/user-attachments/assets/02debfe3-7da5-47e3-8720-d70cf3aee802) ## Dataset Overview ![image](https://github.com/user-attachments/assets/14901599-cd7b-45ba-924b-846bf46df31d) You can preview more [here](simgas_preview.md). ## Pre-computed Masks `result_full` has all pre-computed masks and animated webp images for SimGas, and `gasvid_res_full` has all pre-computed masks and webp for GasVid. You can preview these animated webp: [SimGas](simgas_preview.md) [GasVid](gasvid_preview.md). ## Results | BGS | VLM Filtering | Temporal Filtering | Seg. | τ₍VLM₎ | Stationary Foreground I/P/R/FLA | Moving Foreground I/P/R/FLA | Overall I/P/R/FLA | |------------|----------------|---------------------|--------|--------|-------------------------------|-----------------------------|----------------------------| | ✓ | | | None | - | 0.56 / 0.64 / 0.83 / 0.85 | 0.38 / 0.53 / 0.58 / 0.69 | 0.5 / 0.61 / 0.73 / 0.79 | | ✓ | ✓ | | SAM 2 | 0.09 | 0.67 / 0.81 / 0.79 / **0.88** | 0.54 / 0.79 / 0.65 / 0.83 | 0.62 / 0.80 / 0.74 / 0.86 | | | ✓ | ✓ | SAM 2 | 0.19 | 0.22 / 0.39 / 0.28 / 0.57 | 0.46 / 0.65 / 0.59 / 0.74 | 0.31 / 0.49 / 0.4 / 0.63 | | ✓ | ✓ | ✓ | Trad. | 0.12 | 0.57 / **0.85** / 0.65 / 0.83 | 0.35 / **0.88** / 0.37 / 0.72 | 0.49 / **0.86** / 0.55 / 0.79 | | ✓ | ✓ | ✓ | SAM 2 | 0.12 | **0.70** / 0.83 / **0.82** / 0.87 | **0.69** / 0.79 / **0.84** / **0.92** | **0.69** / 0.82 / **0.82** / **0.89** | **Table: Ablation study of different components with IoU (I), Precision (P), Recall (R), and frame-level accuracy (FLA).** In the segmentation column (Seg.), traditional (Trad.) means Otsu [Otsu] combined with morphological transformations. This analysis corresponds to our ablation study, detailed in Section 4 of the paper. Warning: Different methods of calculating IoU can produce inconsistent results. We used per video aggregation and then average across all videos. ## Test on SimGas Clone the GitHub repo (not this repo) https://github.com/weathon/Lang-Gas/ ### Step 1 Install required packages Install pytorch according to https://pytorch.org/get-started/locally/ ```bash pip3 install opencv-python transformers tqdm Pillow wandb matplotlib scipy ``` Then, run ```bash pip3 install bitesandbytes 'accelerate>=0.26.0' ``` ### Step 2: Download and Install SAM-2 ```bash git clone https://github.com/facebookresearch/sam2.git mv sam2 .sam2 cd .sam2 pip3 install -e . ``` If installation fails, run: ```bash echo -e '[build-system]\nrequires = [\n "setuptools>=62.3.0,<75.9",\n "torch>=2.5.1",\n ]\nbuild-backend = "setuptools.build_meta"' > pyproject.toml ``` (See https://github.com/facebookresearch/sam2/issues/611 for more) Then run: ```bash pip3 install -e . ``` Download the checkpoints: ```bash cd checkpoints bash download_ckpts.sh ``` More details: https://github.com/facebookresearch/sam2 ### Step 3. Download Dataset #### Google Drive Go to https://forms.gle/dJqHdiEN5u8gbVT98 and download the dataset. Put the videos of the downloaded dataset into the `simulated_gas` folder, and run `python3 dataprep.py`. Remove video 24, 26, 27, 28 as needed. (see the paper) #### HuggingFace If you want to use this dataset with HuggingFace, you can download it from [here](https://huggingface.co/datasets/weathon/sim_gas). However, our demo is not integrated with HF, but you can use it in your own project. The HF dataset does not include video 24 to ensure same length for input and groundtruth. We did not provide train-test split as our method is zero-shot. when spliting the dataset, we strongly to split based on video instead of frames. ### Step 4. Run the code #### Modify `owl_notracking.py` for SAM-2 path. Change `sam2_checkpoint = "../../.sam2/checkpoints/sam2.1_hiera_small.pt"` to your SAM-2 checkpoint path. Do NOT change config path. #### Run the code Run `python3 owl_notracking.py` with the following options. If you used `dataprep.py`, the root path should be `sim`. ```bash Usage: owl_notracking.py [OPTIONS] Options: --video_id TEXT Video ID, such as vid20 --root_path TEXT Root path to dataset (default: "../sim") --display TEXT Display server for visualization (default: "localhost:10.0") --log_file TEXT Output log file name (default: "results.csv") --temporal_filter Enable temporal filtering (flag) --vlm_threashold FLOAT Threshold for VLM decision (default: 0.12) --positive_prompt TEXT Positive prompt for VLM analysis (default: "white steam") ``` To reproduce our results, use `--temporal_filter` flag ## Test on GasVid ### Preparation 1. Download the dataset from the original author: [https://drive.google.com/drive/folders/1JKEMtCGPSq2IqGk4uXZP9A6gr8fyJsGC](https://drive.google.com/drive/folders/1JKEMtCGPSq2IqGk4uXZP9A6gr8fyJsGC) 2. Remove all videos captured at depths greater than 18 meters as specified in the original paper. 3. Place the remaining videos in the `Videos` directory. 4. Run `convert_gasvid.sh` to extract frames. ### Running the Test 1. Edit `run_gasvid.py` and `owl_gasvid.py` to set the correct dataset path. 2. To evaluate the entire dataset, run: `python3 run_gasvid.py` 3. To evaluate a single video, run: `python3 owl_gasvid.py --video_id [video id]` 4. Output results are stored in the `gasvid_res_full` directory.