This repository contains an implementation of the (Tiny) Yolo V1 objector detector for both Android and iOS.

This Ted Talk by the creator of Yolo itself gives a nice high-level overview: Joseph Redmon - How a computer learns to recognize objects instantly.

You should also check out his CV. Really, do it ;)

The notebooks provided in this repository contain some more references regarding Yolo.

This project came to be because I wanted to apply the knowledge I have gained from various Deep Learning related courses over the year in a practical project and was searching for a workflow which supports:

• Model Exploration/Implementation
• Model Training/Validation
• Model Optimization
• Deployment on iOS and Android

• Realtime object detection
• Support for Android and iOS
• "Live" Switching between Portrait and Landscape Orientation

The notebooks should be compatible with Python 3.5, Keras 2, Tensorflow 1.2.x. You can find the complete list of dependencies in environment.yml

The Android app is written in Kotlin and should work with any Android Studio Version from 3.x onwards.

Run pod install to install the required dependencies via Cocoapods. The iOS app is written in Swift 3 and Object C++ and should work with a recent version of Xcode.

The notebooks will use this folder to export models to.

This notebook documents the process of implementing Yolo in Keras, converting the pretrained darknet weights for keras and converting them to a format compatible with the tensorflow backend.

This notebook shows how to export the keras model to tensorflow and how to optimize it for inference. The resulting model files contain the tensorflow model that will be loaded by the mobile apps.

iOS: Open the project in XCode and drag and drop frozen_yolo.pb into XCode.

Android: Create a folder named mobile/Android/YoloTinyV1Tensorflow/app/src/main/assets and copy optimized_yolo.pb into it.

Note: You could try to use optimized_yolo.pb with iOS as well. It didn't work with the version of tensorflow I was using though.

The mobile apps do not use Non-Maximum Suppression yet. This means that the apps will display multiple boxes for the same object. I will add this feature to the apps soon. Check out notebooks/01_exploration.ipynb if you're interested in how this works, or you want to implement it youself.

Performance on Android and iOS is suboptimal. There are some opportunities to improve performance (e.g. weight quantization). Will definitely look into this some more.

Both apps only use the back camera. A camera switcher would be a nice improvement.