Practical overview
When it fits
Use RobustSAM when a computer-vision workflow needs promptable segmentation on images affected by degradation, or when researching model training and evaluation for that setting.
Who it is for
- Computer-vision researchers evaluating segmentation on degraded imagery
- Developers building image-processing research workflows that need segmentation
- Teams investigating SAM-based downstream tasks such as dehazing or deblurring
Documented uses
- Segment low-quality images using box or point prompts
- Evaluate segmentation behavior on degraded-image datasets
- Train the model from scratch or from a pretrained checkpoint
- Explore segmentation as part of image dehazing or deblurring research
Requirements & limits
The repository is public and licensed under MIT. The documented local setup requires Python 3.10, Conda, PyTorch, CUDA-specific installation steps and a separately downloaded model checkpoint.
- The project is a model and code repository, not a hosted inference API.
- Inference examples support box and point prompts and include bundled demo images.
- Training guidance includes dataset download and degradation-generation steps; distributed training is documented with a multi-GPU command.
- The repository links three checkpoint sizes: ViT-B, ViT-L and ViT-H.
Check current requirements in the official documentation. Source-based guidance does not mean OpenViST has run or independently tested this project.
Set up RobustSAM for local inference
Prepare the repository, dependencies and a pretrained checkpoint for the documented demo workflow.
Before you start
- Conda and Python 3.10
- A CUDA environment and an NVIDIA GPU, as indicated by the CUDA-version check and PyTorch installation instructions
- A downloaded RobustSAM checkpoint
Steps
Create and activate a Conda environment with Python 3.10.
Clone the RobustSAM repository and enter its directory.
Check the CUDA version with nvidia-smi and install PyTorch, torchvision and torchaudio using the corresponding CUDA wheel index.
Install the remaining dependencies with pip install -r requirements.txt.
Download a ViT-B, ViT-L or ViT-H checkpoint and place it in the repository directory.
Expected outcome
The documented environment and checkpoint are in place for the repository's inference or demo commands.
Guide limitations
- The setup instructions require adapting the PyTorch wheel command to the local CUDA version.
- This setup does not itself run inference; the repository provides separate evaluation commands and demo images.
Run the bundled prompt-based demo
Use the repository's demo images with its documented box- or point-prompt evaluation command.
Before you start
- A prepared RobustSAM environment and checkpoint
- The repository's demo_images folder
Steps
From the repository, run python eval.py --bbox --model_size l for box-prompt evaluation.
For point-prompt evaluation, run python eval.py --model_size l.
Check the documented demo_result directory for the saved results.
Expected outcome
The README specifies evaluation with either prompt type and a default output location under demo_result.
Guide limitations
- The example commands use model size l.
- The README describes the commands and output path but does not guarantee results for every environment or image.
Sources & review
Source record details
- Source availability
- Active
- Sources checked
- 14 October 2026
- Original website title
- robustsam/RobustSAM · Official homepage title · retained in its source language