DART: Detect Anything in Real Time via Segment Anything
Primary Source: https://github.com/mkturkcan/DART
Author: M. Kemal Turkcan · Year: 2026 · Publisher: M. Kemal Turkcan
Summary
A training-free distillation framework that converts heavy foundation segmentation models (SAM3) into a real-time, multi-class open-vocabulary detector achieving 55.8 AP on COCO val2017.
Key Points
- Zero retraining required: distills promptable foundation representations into real-time bounding box detectors.
- Open-vocabulary detection enables finding arbitrary user-specified object classes on the fly.
- Operates at 60+ FPS on edge GPUs and robotics platforms.