Practical overview
When it fits
Use it when you need a locally run research model and have an environment compatible with its documented Linux, Python 2.7, CUDA/cuDNN, and PyTorch 0.3.1 requirements.
Who it is for
- Computer-vision researchers studying image restoration and visual attention.
- Developers who want to run raindrop-removal predictions on local image collections.
- Researchers evaluating restorations against aligned image and ground-truth pairs.
Documented uses
- Generate restored versions of images obscured by raindrops on glass or a camera lens.
- Run model predictions over images stored in local input and output directories.
- Evaluate predictions quantitatively against paired ground-truth images.
Requirements & limits
The repository is public on GitHub and the README documents local installation and command-line prediction and testing.
- The README targets an older software environment: Python 2.7 and PyTorch 0.3.1.
- The listed prerequisites include an NVIDIA GPU, CUDA, and cuDNN.
- Quantitative evaluation uses paired images, and the README identifies test_a as the better-aligned evaluation subset.
- Raindrop-obscured regions may have lost scene information, so their restoration is generated rather than a guaranteed reconstruction of the original.
Check current requirements in the official documentation. Source-based guidance does not mean OpenViST has run or independently tested this project.
Run demo prediction on local images
Use the documented demo mode to process an input image directory and save restored images to an output directory.
Before you start
- Linux
- Python 2.7
- NVIDIA GPU with CUDA and cuDNN; the README specifies CUDA 8.0
- PyTorch 0.3.1 and its dependencies
- A cloned DeRaindrop repository and input images
Steps
Clone the repository and install PyTorch and the project dependencies.
Place images in an input directory, following the README's example of ./demo/input/.
Run
CUDA_VISIBLE_DEVICES=gpu_id python predict.py --mode demo --input_dir ./demo/input/ --output_dir ./demo/output/, substituting the GPU identifier and desired directories.Inspect the output directory for restored images.
Expected outcome
The model writes predicted restorations to the specified output directory.
Guide limitations
- The documented environment is an older stack: Python 2.7 and PyTorch 0.3.1, with CUDA 8.0 listed.
- Restoration of occluded content is a model output; the original scene information may be lost behind raindrops.
Evaluate predictions against paired ground truth
Run the documented quantitative test mode on the aligned test_a images and their ground-truth pairs.
Before you start
- The repository's documented Linux, Python 2.7, NVIDIA GPU with CUDA/cuDNN, and PyTorch 0.3.1 setup
- The test_a dataset arranged under the repository with data and gt directories
Steps
Place the test_a dataset under the DeRaindrop directory, with images in ./test_a/data/ and corresponding ground truth in ./test_a/gt/.
Run
CUDA_VISIBLE_DEVICES=gpu_id python predict.py --mode test --input_dir ./test_a/data/ --gt_dir ./test_a/gt/, substituting the GPU identifier.Review the evaluation result produced by the test workflow.
Expected outcome
The test mode evaluates predictions using paired input and ground-truth images.
Guide limitations
- The README specifies test_a for quantitative evaluation because its image pairs have good alignment.
- The README reports a 0.06 PSNR decrease from the paper result due to masking some information, such as vehicle license plates.
Sources & review
Source record details
- Source availability
- Active
- Sources checked
- 18 October 2026
- Original website title
- rui1996/DeRaindrop · Official homepage title · retained in its source language