mkturkcan / DART

What is DART?

DART transforms SAM3 into a real-time, open-vocabulary object detector—no training required. Feed it text prompts like "person car bicycle" via CLI, and it spits out boxes, scores, and masks on images or videos at 15+ FPS on an RTX 4080 (55.8 AP on COCO with 80 classes). Built in Python with TensorRT exports and ByteTrack tracking, it's a plug-and-play upgrade for "detect anything" demos on this detect anything github repo.

Why is it gaining traction?

It squeezes real-time speeds from frontier models like SAM3 using batched FP16 inference, text caching, and distilled backbones (3-5x faster), without accuracy hacks. Demos run out-of-the-box—`demo_video.py --video input.mp4 --classes car person --trt engine.engine`—and benchmarks compare modes transparently. Devs grab it for quick OV detection prototypes over slower alternatives.

Who should use this?

CV engineers prototyping real-time detection in robotics, dashcams, or surveillance feeds. Video analysts needing tracked objects across frames without custom training. Teams evaluating "detect anything in real time" for apps like AR overlays or autonomous nav.

Verdict

Grab it if you need fast SAM3 detection today—docs, scripts, and COCO evals are solid for a 91-star Python project. Low 1.0% credibility score flags early maturity (rebuild TRT engines per GPU), but it's battle-tested on 4080s; test your hardware first.

(187 words)