Detecting Pikachu on Android using Tensorflow Object Detection
summary
Builds a Pikachu detector with TensorFlow Object Detection, converts labeled XML to TFRecord and trains an SSD + MobileNet pipeline on ~230 images with an 80/20 split then extends the set and trains longer, producing a ~20 MB frozen graph exported with export_inference_graph. Runs the demo on Android via the TensorFlow Android repo after setting optimize_graph=false to avoid missing nodes. Reports workable real‑time detections, many false positives, tunable confidence threshold, and performance vs accuracy tradeoffs.
AI-written summary · about
chapters (18)
- 0:00Introduction
- 1:50Object detection importance
- 3:54TensorFlow API overview
- 4:58Setting up environment
- 6:24Pikachu dataset creation
- 8:09Data to TFRecord
- 8:31Training pipeline config
- 10:15SSD architecture
- 11:18Training and evaluation
- 13:02Freezing model
- 14:30Static image detection
- 17:00Android deployment
- 18:46Android compatibility fixes
- 21:59Improving accuracy
- 22:20Video stream detection
- 24:09Results and limits
- 24:51Future improvements
- 25:53Q&A
related talks
46mTensorFlow for Mobile Poets
40mApplied TensorFlow in Android apps
19mOn-device machine learning: TensorFlow on Android
31mSupercharging Android Apps with Tensorflow
38mDeep Learning for detection on a phone
30mTensorFlow for Mobile Heroes
48mImage Recognition and Object Detection with MediaPipe on A…
43mTensorFlow for Mobile Developers