TinySeg: Model Optimizing Framework for Image Segmentation on Tiny Embedded Systems
summary
Framework called TinySeg implements dynamic tensor spilling and fused fetching on top of TensorFlow Lite for Microcontrollers, targeting an image‑segmentation U-Net variant with ~100k parameters on an Arduino Vision board (dual cores, ~1 MB RAM). It analyzes tensor lifetimes, spills long‑living tensors to local/external flash or host, compresses spilled tensors, uses async block erase, and reports a 39% peak‑memory reduction with only ~6% runtime overhead after optimizations.
AI-written summary · about
chapters (9)
- 0:00TinySeg framework intro
- 1:51Memory challenges
- 2:55Optimization blockers
- 4:00Dynamic tensor spilling
- 5:06Framework components
- 7:14Runtime optimizations
- 9:03Evaluation setup
- 11:15Conclusion
- 12:00Q&A
related talks
45mTech Tutorial: Image segmentation
21mThings I learned running neural networks on microcontrollers
27mMachine Learning on the edge using TensorFlow Lite
19mWhy Machine Learning on Microcontrollers?
27mWhat’s TinyML good for
39mLeveraging the power of C++ for efficient machine learning on embedded devices
55mBuilding with TensorFlow Lite for microcontrollers
01h 53mTinyML: Machine Learning’s Future is Tiny and Bright