Image Recognition, Object Detection, Model Evaluation
summary
MediaPipe on Android uses TensorFlow Lite models (examples: EfficientDet Lite0, EfficientDet Lite2, SSD MobileNet V2) and exposes object detector, image classifier and segmenter APIs with BaseOptions for CPU/GPU selection. Typical detections show confidences around 50–70% (one crow classifier gives 71%), an example IOU reaches ~80–90%, and a hair-specific segmentation model produces tight masks while DeepLabV3 is slower and fuzzier. Model customization uses MediaPipe Model Maker and COCO-format data for training and export.
AI-written summary · about
chapters (14)
- 0:00MediaPipe & on-device ML
- 3:17What is MediaPipe
- 6:50MediaPipe tasks & graphs
- 11:06Vision tasks overview
- 14:17Object detection in Android
- 21:34Configuring object detection
- 29:39Image classification
- 33:51Image segmentation basics
- 37:41Choosing segmentation models
- 40:29Pre-trained MediaPipe models
- 41:34Why on-device machine learning matters
- 43:01Model accuracy & performance
- 45:50Customizing with Model Maker
- 48:18Summary
related talks
12mCascade models in computer vision to boost accuracy and performance
26mWhere can I park? Deep Learning is here to help you!
45mTech Tutorial: Image segmentation
55mCore ML: Vision
36mDeep Learning - Segmentation, Identification and Classification
28mObject Detection using Transformers and CNNs
21mLeveraging Python for Real-Time Image processing
37mSeeing is Believing: Mobile Vision API Deep Dive