Deep Learning for detection on a phone
summary
Pipeline shows how a detector gets from prototype to phone: configuration-driven deployment file, model definition converted to convolutional-only layers, training with transfer learning, and early profiling on device using Metal Performance Shaders and Caffe bindings. Targets: real-time 60 FPS with tiny footprints like ~5 MB, 1–2M parameters. Uses public corpora COCO and Pascal and KITTI, extensive relabeling, precision–recall evaluation, frozen early layers, and automated experiment tracking to decide what actually works.
AI-written summary · about
chapters (11)
- 0:00Deep learning for phone detection
- 2:14Why deep learning was needed
- 3:16Full pipeline for phone deployment
- 6:49Choosing mobile model architecture
- 11:32Dataset selection and prep
- 15:07Training with transfer learning
- 18:40Evaluating beyond metrics
- 22:14Deploying on mobile devices
- 24:04Data and evaluation pitfalls
- 27:14Summary
- 27:56Q&A
related talks
40mDeep Learning on iOS
58mDeep Learning in Real Life
47mDeep Learning: a practical introduction from an Android dev perspective
12mCascade models in computer vision to boost accuracy and performance
59mDeep Neural Networks for Video Applications
32mApplying Deep Learning for Object Detection from a Telecom Perspective
46mFire Up Your Smartphone Neurons with On-Device Deep Learning
27mMachine Learning and Mobile