PySIFT – GPU-Resident SIFT for Computer Vision

PySIFT is a pure-Python, open-source GPU-resident implementation of the Scale-Invariant Feature Transform (SIFT) built on CuPy and Numba CUDA kernels which is faster and yet more accurate. It runs the entire detection-to-descriptor pipeline on your NVIDIA GPU with zero-copy DLPack interop to PyTorch so that your downstream DL steps will be free from CPU PCIe bottlenecks.

PySIFT is the feature extraction foundation for downstream pipelines in medical imaging, drone/UAV stitching, SLAM, robotics, 3D reconstruction (NeRF, 3DGS, SfM), and satellite remote sensing.

pip install staysift

Key Results

Benchmarked against OpenCV SIFT on RTX 3050 Laptop GPU (4 GB VRAM):

Benchmark	Metric	PySIFT	OpenCV	Delta
HPatches	MMA@10	0.703	0.681	+2.2pp
IMC Phototourism	Inliers/pair	303	205	+47%
MegaDepth-1500	AUC@10	0.503	0.447	+5.6pp
ROxford5K	mAP (Medium)	0.455	0.380	+7.5pp

Links

• Paper: arXiv:2605.17869

• Code: github.com/SivaIITM/PySIFT

• PyPI: pip install staysift

• Tutorial: sivaiitm.github.io/PySIFT

• HF Space: huggingface.co/spaces/sivaIITM/PySIFT

• Kaggle: IMC 2026 Warm-Up Sprint

Contents

• Installation
  • Step 1: GPU Dependencies
  • Step 2: Install PySIFT
  • Optional Dependencies
  • Verification
  • Hardware Tested
  • No GPU?
• Quick Start
  • Feature Extraction
  • GPU-Resident Output
  • Feature Matching
  • Panoramic Stitching
  • CLI Stitching
  • Configuration Presets
• API Reference
  • PySIFT
  • GPUPyStitch
  • Zero-Copy DLPack Interop
• Benchmarks – PySIFT vs OpenCV SIFT
  • HPatches – Matching Accuracy
  • IMC Phototourism – Pose Estimation
  • MegaDepth – Wide-Baseline Stereo
  • ROxford5K – Image Retrieval
  • Speed
  • Resolution Scaling
  • Ablation: Classical vs Learned
• Architecture – The 5-Stage GPU Pipeline
  • Stage 1: CLAHE Preprocessing
  • Stage 2: Gaussian Scale-Space Pyramid
  • Stage 3: DoG Extrema Detection
  • Stage 4: Orientation Assignment
  • Stage 5: Descriptor Computation
  • Zero-Copy DLPack Handoff
  • VRAM-Adaptive Execution
• Use Cases – PySIFT in Downstream Pipelines
  • Medical Image Registration
  • Drone / UAV Aerial Stitching
  • SLAM and Visual Odometry
  • Robotics – Object Localization
  • 3D Reconstruction (NeRF / 3DGS / SfM)
  • Satellite and Remote Sensing
  • Comparison: PySIFT vs Learned Detectors for Downstream Use