SSeg: Active Sparse Point-Label Augmentation for Sementic Segmentation

CV4EO Workshop - WACV 2026

Cesar Borja, Carlos Plou, Ruben Martinez-Cantin, Ana C. Murillo

DIIS - Instituto de Investigación en Ingeniería de Aragón (I3A), Universidad de Zaragoza, Spain

🌍 Homepage | 📝 Paper |

---

🔔 News

• 🆕 01/2026: Code released.
• 🥳 01/2026: Paper accepted at Computer Vision for Earth Observation Workshop Series - WACV 2026.

---

📖 Description

SSeg is a framework for efficient multi-class image segmentation that expands sparse point annotations into full masks using a hybrid SAM2–superpixel strategy with active point sampling

🛠️ Installation

PyTorch must match your CUDA driver, so installation is a two-step process (torch first, everything else second). See INSTALL.md for the manual walkthrough and troubleshooting.

Quick start (recommended):

bash setup.sh                            # creates conda env "spuw", auto-detects CUDA
conda activate spuw
bash checkpoints/download_ckpts.sh       # ~900 MB SAM2 checkpoint
python run.py                            # smoke-tests the demo experiment

setup.sh flags: --env-name <name> (default spuw), --cuda <ver> to override detection, or --cpu for a CPU-only install.

Requirements: Linux, NVIDIA driver ≥ 525 with CUDA ≥ 11.8 (CPU works too — just slower), Python 3.10, conda, ~5 GB free disk.

🚀 Usage

You can run the framework in two modes:

Option A: Interactive GUI

Best for testing single images and visualizing results in real-time.

python app.py

Pick a folder of images with Select Folder, then press Start to load the segmenter and begin annotating. The tool suggests the next point to label (yellow cross) and shows a live preview of the mask as you hover.

Controls:

• Left click — place a prompt. Toggle the Negative / Positive button to switch between positive points (include) and negative points (exclude).
• BBox — click the button to arm box-prompt mode, then click two opposite corners to draw a bounding box prompt; a live preview is shown between the two clicks. Points and boxes can be combined for the same mask.
• ✓ (Finish) — expand the current prompts into a mask and assign it a label.
• 👁️ (Toggle masks) — show / hide the committed masks.
• Right click on an existing mask — delete it or change its label.
• Ctrl + mouse wheel — zoom the display in (wheel forward) / out (wheel backward) for finer annotation. Zoom starts at 100% (fit-to-window); when the zoomed image no longer fits, horizontal and vertical scrollbars appear so you can pan. Zooming only affects the display — point and box coordinates stay at full image resolution.
• Ctrl + Z — undo the last point or box (in placement order).
• < / > — move to the previous / next image (prompts to save first).

Option B: Benchmark/Batch Mode

Best for running experiments on full datasets and generating metrics.

python run.py

(Requires ground truth images for point-label assignment)

Configuring Batch Experiments

To configure batch experiments, you do not use command-line arguments. Instead, you modify the configuration list directly inside run.py.

How to Add Experiments

Open run.py and scroll to the experiments list (around line 220). Add a dictionary for each experiment you want to queue:

experiments = [
    {
        "name": "experiment_name",      # Folder name for results
        "strategy": "dynamicPoints",    # Sampling strategy
        "num_points": 25,               # Point budget
        "images": "path/to/images",     # Input directory
        "ground_truth": "path/to/gt",   # GT directory (required for dynamic strategies)
        
        # Advanced Parameters
        "lambda_balance": 0.5,          # (Dynamic only) Balance exploration/exploitation
        "heatmap_fraction": 0.5,        # (Dynamic only) % of points from uncertainty map
        "maskSLIC": True,               # Enable superpixel refinement
        "visualizations": True          # Save debug images
    }
]

Parameter Reference

These keys can be used inside the experiment dictionary in run.py:

Required

• name: Identifier for the experiment (creates output subfolder).
• strategy: Active sampling logic: random, grid, list, dynamicPoints, SAM2_guided.
• images: Path to the directory containing input images.

Optional (General)

• ground_truth: Path to ground truth masks. Required if using dynamic strategies or for evaluation.
• num_points: Total point budget per image (default: 30).
• color_dict: Path to JSON file mapping colors to class IDs (essential for RGB GTs).
• default-background-class-id: Integer ID to use for the background class (default: 0).
• maskSLIC: Set to True to enable MaskSLIC superpixel refinement.
• visualizations: Set to True to save overlay images for debugging.

Optional (Strategy Specific)

• lambda_balance: (Dynamic only) Float [0-1]. Higher values favor exploitation (coverage).
• heatmap_fraction: (Dynamic only) Float [0-1]. Ratio of points sampled from uncertainty heatmap.
• strategy_kwargs: A dictionary of extra parameters specific to the chosen strategy (e.g., {"propagation_overlap_policy": "last"}).

Debugging

• debug_expanded_masks: Set to True to save individual SAM2 expansion steps.

📜 License

Discuss with your supervisor the license that you want to set and include the licenses of any previous repo in which your code was inspired.

📝 Citation

@article{borja2025sparseuwseg,
  title={SparseUWSeg: Active Sparse Point-Label Augmentation for Underwater Semantic Segmentation},
  author={Borja, C{\'e}sar and Plou, Carlos and Martinez-Cant{\'\i}n, Rub{\'e}n and Murillo, Ana C},
  journal={arXiv preprint arXiv:2510.10163},
  year={2025}
}

🙏 Acknowledgements

This work was supported by a DGA scholarship and by DGA project T45_23R, and grants AIA2025-163563-C31, PID2024-159284NB-I00, PID2021-125514NB-I00 and PID2024-158322OB-I00 funded by MCIN/AEI/10.13039/501100011033 and ERDF.