AutoTune

Introduction

This is an implementation of paper Autonomous Learning for Face Recognition in the Wildvia Ambient Wireless Cues

Install

Environment

OpenCV
tensorflow 1.4
Tshark
goprocam

How to run

Preparation

Wifi data caputre

Set your router to make sure the router to have fixed channels
Prepare a computer running ubuntu 14.04
Add mac address which you want to listen to config.yaml file. Example like that:
```
mac_name:
 'device:mac:address': device's name
 ...
```
Run create_listen_bash.py to create bash scripts.
- Firstly, run listen.sh to create interface for listening
- Secondly, run chanhop.sh to listen the fixed channel set in step 1, the data file will be saved in {root_folder}/wifi_data/.
- If you want to stop listening process, kill the two bash above, then run wifi_mod.sh to reset the network interface card.
Tips: Due to the differences of network interface cards, you may need different commends to start the listening process. Although create_listen_bash.py script will create two groups of bash script for different devices, it is very likely that none of them can work on your devices. You may need to try the right commend in your decice. The network interface card we use is TP-Link TL-WN726N

Video Data caputre

To begin with, you will prapare some GoPro Hero 4, setting these device's video quality to 120fps and 720p.
Set these devices to Wi-Fi mode and then connect your computer to the GoPro's Wi-Fi.
Make sure your computer installed goprocam and run live_stream.py to capture video.

Data Preprocess

Wi-Fi Data

Run analysis_wifi.py to preprocess the Wi-Fi data. the result will be saved in {root_folder}/origin_data/

Video Data

Run download_video_from_gopro.py to download these videos captured by gopro, these videos will be saved in {root_folder}/video/ and be named by those videos' capture time.
Run assign_video_by_shot_time.py to group videos based on capture time, these videos will be moved to {root_folder}/origin_data/.

Preprocess

Run source/pre_process.py to convert videos to aligned face images and the result will be saved in {root_folder}/middle_data/time/.
Run source/get_meeting_data.py to assgin data based on session setting.

Label Data

During tuning process, Autotune framework will evaluate the framework performance iteratively.

Runing source/feature_extraction_classifier.py

After running, the structure of result folder is showing below. You need to correct the classifier error in the classifier folder manually.

middle_data/
├── 03-22-11-00-00_03-22-14-00-00
│   ├── classifier
│   │   ├── 0 # people 0's images
│   │   │   ├── 03-22-12-20-00_03-22-12-40-00_4787_1079.jpeg
│   │   │   ├── 03-22-12-40-00_03-22-13-00-00_134_2127.jpeg
│   │   │   ├── 03-22-12-40-00_03-22-13-00-00_135_2251.jpeg
│   │   │   ├── 03-22-12-40-00_03-22-13-00-00_136_2079.jpeg
│   │   │   └── 03-22-12-40-00_03-22-13-00-00_137_2389.jpeg
│   │   └── 1 # people 1's images
│   │       ├── 03-22-12-40-00_03-22-13-00-00_139_2337.jpeg
│   │       ├── 03-22-12-40-00_03-22-13-00-00_140_2317.jpeg
│   │       ├── 03-22-12-40-00_03-22-13-00-00_141_2300.jpeg
│   │       ├── 03-22-13-40-00_03-22-14-00-00_4124_7513.jpeg
│   │       └── 03-22-13-40-00_03-22-14-00-00_4125_6286.jpeg
│   └── mtcnn # origin data
│       ├── 03-22-12-20-00_03-22-12-40-00_4787_1079.jpeg
│       ├── 03-22-12-40-00_03-22-13-00-00_134_2127.jpeg
│       ├── 03-22-12-40-00_03-22-13-00-00_135_2251.jpeg
│       ├── 03-22-12-40-00_03-22-13-00-00_136_2079.jpeg
│       ├── 03-22-12-40-00_03-22-13-00-00_137_2389.jpeg
│       ├── 03-22-12-40-00_03-22-13-00-00_139_2337.jpeg
│       ├── 03-22-12-40-00_03-22-13-00-00_140_2317.jpeg
│       ├── 03-22-12-40-00_03-22-13-00-00_141_2300.jpeg
│       ├── 03-22-13-40-00_03-22-14-00-00_4124_7513.jpeg
│       └── 03-22-13-40-00_03-22-14-00-00_4125_6286.jpeg
└── 03-22-14-00-00_03-22-17-00-00
    ...

AutoTune Hard

cd source
./auto_em_hard.sh

AutoTune Soft

cd source
./auto_em.sh

ml-lab / autotune Goto Github PK

autotune's Introduction