Reproduce “You Only Look Once at Anytime (AnytimeYOLO): Analysis and Optimization of Early-Exits for Object-Detection” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Daniel Kuhse, Harun Teper, Sebastian Buschjäger, Chien-Yao Wang, Jian-Jia Chen
This research introduces AnytimeYOLO, an adaptation of the YOLO architecture that integrates early-exits to enable interruptible object detection, allowing for dynamic quality-latency trade-offs in real-time applications. A novel transposed architecture (GELAN-T) notably improves early prediction quality, while graph-based optimization algorithms enhance deployment efficiency by determining optimal execution paths.
Key insights
- Integrating early-exits into YOLO architectures, by repurposing deep supervision, enables fine-grained interruptible object detection, offering adaptability in real-time systems.
- Architectural modifications, such as the transposed GELAN-T design, can significantly improve the quality of early predictions by making multi-resolution features available sooner, though sometimes at a slight cost to final accuracy.
- Optimal path selection through graph-theory-based algorithms effectively optimizes anytime model deployment by finding superior execution orders and exit subsets compared to greedy heuristics.
- Current GPU inference frameworks present practical challenges for true "hard" anytime inference due to asynchronous operations and whole-network optimization, leading to non-negligible overheads for early-exit retrieval.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.