UltraPoint: improving SuperPoint for faster image keypoint extraction

We present UltraPoint, a lightweight neural network model for keypoint detection and description that improves upon the SuperPoint architecture. UltraPoint uses a streamlined encoder based on depthwise separable convolutions to drastically reduce computation—achieving about a 6× reduction in FLOPs and a 4× reduction in parameters compared to SuperPoint—while preserving the quality of the local features. On the HPatches benchmark, UltraPoint attains repeatability and localization accuracy on par with the original SuperPoint. The model is significantly smaller than current state-of-the-art learned feature extractors, enabling real-time performance (60+ FPS on a mid-range GPU) even on devices with limited computational resources. UltraPoint is trained via self-supervised learning (homographic adaptation on real images) and can also leverage knowledge distillation from a pre-trained SuperPoint to retain high descriptor discriminability. The resulting model provides an excellent trade-off between speed and accuracy for tasks like image matching, visual odometry, SLAM, and augmented reality. We also outline future extensions such as incorporating rotation invariance and applying model quantization to further improve efficiency.