Digitize my home movies

Overview

Primarily, I'm trying to back things up. Secondarily, I'd like my family to be able to watch these clips easily at home on their phones or computers. I have a backup drive that's exposed via samba. Since these clips are so large (2GB for a 2 hour movie), I'd like to make them accessible in a segmented format. My first attempt is to get it working with HLS and a single TS asset via byte-range. If that doesn't work, I might switch to fragmented mp4 (dash or HLS/fmp4). If that doesn't work, screw it; I'll just use mp4.

VHS

I bought a cheap composite video/audio USB adapter off amazon:

• http://a.co/a7Opzim

I already had a VCR and my home computer (a mid-range NUC running Ubuntu). I'll continue to serve these clips via read-only samba and I'll probably add a web server with some static-ish pages and a player to serve segmented content.

At first, I was using vlc for encoding, but I switched to ffmpeg.

Video just worked. Well, it presents as 25 fps and 720x576. But ffmpeg seems to not run right on it's own there. I have to manually override the input and output frame rate at 25. Took me a while to figure that out. Even then it wants to use "variable frame rate". Not sure why this is complicated.

• Video Input: f4l2, /dev/video0

Audio on Linux is such a pain. I had to modify my /etc/pulse/default.pa among other non-sensical things. Took a couple hours of fiddling.

• Audio Input: alsa, hw:1,0

8mm

We have a ton of super 8 film. I bought the "Wolverine Pro" conversion device. It generates MP4 clips.

The only issue I had is that my device seems to grab the film a little too tight on the capture tray. When that happens, it stalls and generates duplicate images (sometimes stalling indefinitely until I manually intervene :-( ). The content is also pretty grainy (our cameras weren't very nice), so I down-res from 1440x1080 to 720x540 so it looks a little cleaner.

List of scripts

Tool	Description
auto-capture.pl	Capture video/audio and stop at end of tape
makefile	Build compressed content from captures (after auto-capture.pl)
web/makefile	Build m3u8 manifests and html to view encoded clips

Supporting tools:

• blank-frames.pl - Find blue (the VCR inactive screen) or black (blank video)
• blank-frame.pl - Find blue (the VCR inactive screen) or black (blank video) at a particular location (in seconds)
• hls-byte-range.pl - Given the output of probe.pl, generate an HLS byte-range manifest
• probe.pl - Find key video frames (pts, time and byte offset)
• web/scripts/manifest.py - Create/modify json metadata blob for a clip
• web/scripts/create_video_html.py - Create html for a single clip (will probably go away)
• web/scripts/create_root_html.py - Create html to link all clips

Software Dependencies

The following items are used during capture and encoding:

• ffmpeg
• convert (imagemagick)
• mediainfo

Things that should already be installed on any normal linux-y system:

• perl
• python
• gnu make
• normal unix tools: mv, mkdir, find, file, touch, ln, echo, date, grep, sed

For my web server, using the default config:

• lighttpd

How to: Capture and conversion

To capture content to a new file, capture/$datetime.ts:

    ./auto-capture.pl

When you have some spare cpu cycles (i.e. when you're not capturing), to encode all existing captures to more compressed variants. The new encodes are stored in the content/ directory but have the same name as their sources in capture/.

    make

Note that this process does not (yet) delete captured sources. I've been comparing the capture vs content as I tweak encoding settings.

Todo Items

0 Metadata ** Backup metadata files ** Use multi-line output that's easier to read/edit by hand ** Improve format for search (use a db? sigh)

1 Capture automation ** auto-capture could both start and stop automatically (wait for not blank to start) ** auto-delete capture/ content after transcode completes successfully ** run web/makefile after transcode

2 Web ** Create page that actually looks nice *** Show video inline ** Update metadata *** Update list and frequency of people (Grace:20%, Jacqui:10%, ...) *** Edit/create scenes ("Hoo, like that":[03:20,06:30]; etc.) ** Add search feature for finding things based on metadata content

3 Face recognition ** Use something like https://github.com/ageitgey/face_recognition, or dlib.net directly ** Construct a web page so Jacqui can train it herself (who is this? is this also them?...) Mechanical Turk: My Wife ** Once enough samples exist, process all video and picture files, updating the metadata about who is in them (and how often, for video).

4 Index all photos and videos ** Put all the above together on my entire photo/video library