--- title: RIPE & RIPE++ Keypoint Matching emoji: πŸ“š colorFrom: gray colorTo: indigo sdk: gradio sdk_version: 5.35.0 app_file: app.py pinned: false license: other short_description: Demo of RIPE (ICCV 2025) and RIPE++ (LIMIT @ ECCV 2026). --- # RIPE & RIPE++ Keypoint Matching Demo A Gradio demo for two keypoint extractors from our group, side by side: - **RIPE** β€” *Reinforcement Learning on Unlabeled Image Pairs for Robust Keypoint Extraction* (ICCV 2025). [arXiv](https://arxiv.org/abs/2507.04839) Β· [Project page](https://fraunhoferhhi.github.io/RIPE/) Β· [Code](https://github.com/fraunhoferhhi/RIPE) - **RIPE++** β€” our latest model (LIMIT @ ECCV 2026, MalmΓΆ). [arXiv](https://arxiv.org/abs/2608.19693) Β· [Project page](https://fraunhoferhhi.github.io/RIPEpp/) Β· [Code](https://github.com/fraunhoferhhi/RIPEpp) Pick a method, upload an image pair (or choose an example), and the demo: 1. extracts the top-2048 keypoints from each image, 2. matches them with the selected matcher, 3. filters outliers with RANSAC (fundamental matrix) when more than 8 matches are found, 4. visualizes the detected keypoints and the raw / RANSAC-filtered correspondences. Images are resized so the smaller side lies within [512, 768] px, aspect ratio preserved. ## Matchers - **Nearest Neighbor** β€” mutual nearest-neighbor matching (kornia MNN for RIPE; the glue-factory NN matcher for RIPE++). - **LightGlue** β€” our trained LightGlue matcher (from glue-factory). **Available for RIPE++ only.** Both models are trained **without geometric ground truth** (no depth, pose, or annotated correspondences) β€” only image pairs and a geometric reward. ## How it's built The method source is **vendored** directly in this repo β€” `ripe/` (RIPE) and `ripepp/` + `gluefactory/` (RIPE++) β€” so nothing is installed from git at build time. Model weights are vendored too: - `weights_ripe.pth` β€” RIPE. - `weights/ripe++.ckpt` β€” RIPE++ extractor. - `weights/ripe++lightglue.tar` β€” RIPE++ LightGlue matcher. The app is decorated with `@spaces.GPU`, so it runs on ZeroGPU: the GPU is attached only for the duration of each inference call. ## Licenses - `ripe/` and `weights_ripe.pth` β€” RIPE, Fraunhofer academic-use license (see the repo-root `LICENSE`). - `ripepp/` (+ RIPE++ weights) β€” RIPE++, Fraunhofer academic-use license (`ripepp/LICENSE`). - `gluefactory/` β€” vendored from the [glue-factory](https://github.com/JohannesK14/glue-factory) fork (based on [cvg/glue-factory](https://github.com/cvg/glue-factory)), **Apache License 2.0** (`gluefactory/LICENSE`). Everything else is installed from PyPI (see `requirements.txt`) under its own license. ## Running locally ```bash pip install -r requirements.txt python app.py ```