Detect Anything in Real Time: From Single-Prompt Segmentation to Multi-Class Detection

Dimensions overall score 9.0

GitHub Code Pulse

No public code linked for this paper yet.

• DART achieves a 5.6x cumulative speedup at 3 classes and scales to 25x at 80 classes compared to SAM3's single-prompt processing.result 95%
• DART reduces the computational cost of the visual backbone from O(N) to O(1) by sharing backbone computation across all classes.technical 95%
• On COCO val2017, DART achieves 55.8 AP at 15.8 FPS (4 classes, 1008x1008) on a single RTX 4080.result 95%
• DART surpasses purpose-built open-vocabulary detectors trained on millions of box annotations in terms of performance.result 90%
• SAM3 processes only a single text prompt per forward pass, requiring N independent executions for N categories.limitation 95%
• DART is a training-free framework that converts SAM3 into a real-time multi-class detector without modifying any model weight.method 95%
• For extreme latency targets, adapter distillation with a frozen encoder-decoder achieves 38.7 AP with a 13.9 ms backbone.result 90%
• DART exploits the structural invariant that SAM3's visual backbone is class-agnostic, producing image features independent of the text prompt.technical 90%

Competitive landscape

Competitor map is still being generated for this paper. Enable generation or check back soon.

Startup potential card