Timeline is a web application which aims to organize photos similar to how Google Photos or Amazon Photos does. Instead of Google or Amazon Photos it is not running in the cloud but on you own Device (Server, NAS, PC)
- Timeline reads and monitors photos from a specified directory on your device. It doesn't touch any of your photos, all data is written elsewhere
- It will display photos in a timeline based view, ordered by date (newest photos first) with fast endless scrolling functionality (no paging)
- Timeline extracts exif data. A limited set is visible in the information panel
- It will detect faces on the photos using MTCNN (Tensorflow)
- Once some faces are detected, they can be named.
- Other faces are matched against the known faces. This works quite well for most faces except for infant and baby faces where it is not very good in separating them
- In order to ease up the face classification it clusters similar faces for easy naming
- It tries also to do some object detection in the photos using Tensorflow; currently using
- If GPS information is present in the Exif Data it will do a reverse geo location look to resolve the address
- When viewing an Photo in full screen you will see of of the detected information (Identified and non identified faces, things a resolved address along with a little map and some important Exif data)
- now also playing HEIC, MOV, and MP4 files
- Selection of Photos via left and right cursor keys
- Rating of Photos via the usual Rating stars either by clicking the rating while hovering over a photo or selecting a photo via cursor left/right and use the 0-5 keys (filtering for the so rated photos will follow)
- Option to change the preview size to accommodate more or less on the photo wall
- Selection of Multiple Photos at one by Shift-Click
- Albums can be created (and deleted)
- New Albums Overview
- Photos can be removed from the Catalog
- New Search View. Photos can be searched by Date Range, Rating, Camera Make, Country, City, People
- A new default Album "Last Import" is created and filled by last added photos accessible with a new navigation shortcut
- A new simple view showing the photos to be sorted into the index is available
- Major re-arrangement with regards to workers. Now only one worker process is required - should be a bit faster but due to loading the models now in each preforked worker process they consume more memory. On something with only 4GB RAM the number of worker should only be 1. With 8GB 3-4 should be ok. Due to using preforked workers they will now be scaled down when not needed anymore.
- Searching for Photos by date, rating, location can now be saved as "smart album" (not really smart but I didn't find a better name). Once defined they also can be edited afterwards.
- For faces it is now detect the expression (happy, sad, neutral ...). This will be later used to find some nice photos of happy people. I was also trying this with gender and age but it was not very precise
- Sectioning is now happening already while importing photos. With this new photos are visible immediately. The background sectioning still happens and is used to equally size the sections in terms of the contained photos
- Uploading of Phots is now possible. Therefore a folder has to be specified which is created below PHOTO_PATH. Photos will be uploaded there and sorted by year and month
- Now also reading and displaying (iPhone) HEIC files
- Addionally also reading and displaying MOV and MP4 files. MOV files will be converted to MP4, so both types will be playable in the Browser. Similar to Google Photos a preview of the video files will played back when hovering over a video
timeline_video.mp4
- Frontend reworked in order to use the space more effectively. Again trying to mimic Google Photos for the layout a bit more
- Separated a dedicated worker for video conversion. Otherwise ffmpeg will eat all CPU resources. Now, there is only one video conversion task happening at a time
- Fullscreen videos will now be transcoded on demand (when clicked). A spinner indicates the conversion. This behaviour can be configured (VIDEO_TRANSCODE_ON_DEMAND in env-File)
- The basepath of the Web Application is not / anymore but configurable path. Default is /timeline. The reason for a basepath is to be able to put this behind one other webserver (nginx, caddy) so that it allows it to proxy different applications under one domain (e.g. some dynamic dns with nextcloud, darktable ...)
- Experimental: When all assets are processed, Timeline attempts to find events. These events are identified by a certain amount of photos within a certain timeframe. The is again configurable by EVENT_MIN_SAMPLES and EVENT_HOURS_EPSILON. Again the DBSAN Algorithm is utilised for this. So identified Events will be created as Albums. The Album name will be constructed of a Date and a Location name (City, County ...). The Presentation of Albums clearly needs to be improved.
- Bug fixing
- Finding co-occurrences of people over time and creating albums of of it
Please note that it detects more than just one thing.
There is also a Dark Mode and an information panel showing what is happening in the background.
There is also a new Seach View (to be extended) to query photos by various combined attributes
Photos can be selected and added to existing or new Albums
The easiest way is to use Docker, resp. docker-compose
-
wget https://raw.githubusercontent.com/hulk66/timeline/main/docker-compose.yml -
wget https://raw.githubusercontent.com/hulk66/timeline/main/env, modify the env file according to your needs and save it as.env. The following things have to be changed:ASSET_PATH- this points to the directory where your assets are located. Timeline will recursively scan all Photos and Videos (JPG, HEIC, MOV, MP4). This is read-only: Timeline will not change anything in here.PREVIEW_PATH- For performance reasons this directory is used to generate all kinds of previews.LOG_PATH- This is where Timeline does the logging. Most important here is theprocess_worker.log.DATABASE_DATA- Here is where all data from mariaDB is savedRABBITMQ_DATA- Same for RabbitMQ which is used as the message brokerWORKERS_PROCESS- Number of workers to be used. This depends on the machine you are running Timeline on. For everything below 8GB RAM only 2 or 3 workers are recommended. If you have more RAM then also 4-6 might be okDB_SUPER_USER_PW- is the password for to access the database (via adminer)- If you want a different basepath then
timelinein the URL this can be configured as well. For the timebeing it has to be some path and must not be empty
If used with Docker please specify all paths absolute.
Important: All directories have to exist. They won't be created.
Once you have the .env file in place run it with
docker-compose up or docker-compose up -d to run it in the background.
Given everything is downloaded from Dockerhub (mariaDB, RabbitMQ ...) it may take 3-4 minutes to start up all containers depending on the system.
If all containers are up and running you should be able to see some change in the worker.log in the specified LOG_PATH.
- To access Timeline go to
http://<whereever_it_runs>:9090/timeline. - In case you are interested in some internals you can also go to
http://<whereever_it_runs>:9091to reach adminer and have a look into the underlying database. The user is "root" and the password is the one you have specified in the.env(example) - There is also running the user interface for RabbitMQ under
http://<whereever_it_runs>:15672. Username and Password is per default guest/guest
When running for the first time give it some time to start and process initially some photos. Timeline is doing a "sectioning" of photos. Photos are sorted chronological in sections. A section is a set of photos (300-500) that are chronological next to each other. This section information is then used on the frontend side in order order to load the related photos. More importantly a section of photos is also unloaded if not visible anymore. Otherwise the browser is getting slower once it comes close to 1000 photos to be displayed.
The sectioning is configured to happen every 15 minutes.
There are two other background tasks:
- Once the user starts to name Faces, Timeline will try to match similar faces which have not been identified yet. This will happen every 30 minutes.
- Every 45 minutes Timeline will try to do face clustering for unidentified faces. Similar faces are grouped together for easy naming in the People view. If there is a new cluster, it will be indicated with a small blue badge on the People Icon.
Timeline was tested in Docker instance with limited resources: 4GB of RAM and 2 Cores might work. Don't try it with less.
As always: the more the better but 8GB RAM and 4 Cores is recommended.
While the user interface should be reasonable fast the AI based face and thing recognition are really slow. User patience is required. On a decent NAS with a AMD RX-421ND (Quad-Core 2,1 GHz) processor and enough RAM (16GB) Timeline processes approx. 500 Photos per hour including face, thing recognition and GPS reverse lookup. For the geo lookup OpenStreepMap is used. In order not get a "Too Many Requests" error the the request rate is limited to 1 request per second. A fast laptop (Intel i7, 6 Cores, 32GB) will process around 2000 photos per hour.
Timeline is heavily using a number of great other tools and frameworks. Just to list the most prominent ones:
Backend
- Python 3.8 as implementation language
- Flask as framework for REST services for the frontend
- SQLAlchemy for doing the object relational mapping
- When using Docker: mariaDB as the underlying relational database
- Celery as task queueing system for handling all asynchronous photo tasks
- When using Docker: RabbitMQ as message broker used by Celery
- Tensorflow for doing object and thing detection and also for doing all face recognition tasks. The models used for this are MTCNN and VGGFace
- GeoPy for doing the reverse geo location lookup, using OpenStreetMap
Frontend
- VueJS as the elementary frontend framework
- Vuetify for providing all the nice material components
- nginx as the webserver acting as the proxy for the webapp and for efficiently delivering the generated previews
- During the the processing it may happen the the database provide a Timeout error due to a locked resource. This actually should not be a problem as the so failed task is repeated at a later point in time by Celery
Where should I start?
- Remove bugs
- Improve AI tasks (faces and things) performance: not sure I will have the capacity and/or knowledge for this
- Improve user interface, especially implement search capabilities for People, Things and places
- Improve frontend performance by generating previews for Places and Things view
- Implement (Smart) Albums: done
- Implement Ratings. For both point the database model is already prepared: Done
- This is nearly test-free code (which is very bad). Needs to be improved
- Multi Language support. Nothing done here at the moment.
- Ability to upload new photos
- Ability to read also RAW photos
Copyright © 2020-2023 by Tobias Himstedt, Sergii Puliaiev. All rights reserved.
See also GNU General Public License version 3
Please note: this project is a one-man-show for now and I'm doing this next to my work. Don't expect something professional.