A simple Fabric.js-based web app for annotating SVG files

In the check-it-out category, I want to share a web app prototype I wrote recently for a client. It was my first experience with writing non-trivial JavaScript (peepweather.com uses a little - its source is on GitHub at github.com/matthewcornell/peepweather) and was a lot of fun to write. The app had no polish, as you can see from the screen shots, because all they wanted was for me to quickly write a solid starting point for an upcoming fall project of theirs.

Following is some detail from the README on the GitHub fork I created. You can play with it at annotator-demo.herokuapp.com.

Description

This is a simple Flask-based web app that demonstrates using Fabric.js to implement a direct-manipulation SVG annotation tool. It has a bare-bones UI and layout (i.e., no Bootstrap or equivalent) and no production-level features such as authentication, authorization, document owners and assignments, error handling, concurrent access, deployment, etc. However, it does support most of the front end features desired, plus the connection to the back end. The original prototype supported adding new svg files by dropping them into the repository directory, which a DAO would pick up, creating corresponding json files as needed. This demo uses a fixed repository of a few files to keep it simple. The point is mainly to show the front end proof-of-concept using Fabric.js.

Motivation

The fuller version of this tool was used to help a group of people annotate thousands of PDF files in SVG format as generated by Mozilla Labs' pdf.js. The SVG retains formatting and so is browser-renderable, but the markup is more useful for information extraction than that of PDF. The ultimate goal of the project was to create a gold standard for comparing the IE output to. Along with saving a document's annotations, users could ask the server for the text bounded by a particular annotation rectangle - a non-trivial problem with these complex SVG files (the 's generated are quite chopped up).

The decision to to write a tool was made after I researched existing annotation tools including some amazing ones:

• Annotator
• ZURB JavaScript Annotation Plugin
• Annotorious
• Annotation Studio

Unfortunately they lacked the customization and power user features we needed for the high-throughput workflow necessary to annotate 1000s of SVG files. I then surveyed JavaScript graphics libraries (both Canvas- and SVG-based) to find one to support the rectangle-based UI we wanted. I settled on the excellent Fabric.js (the demos are awesome) after having looked at general purpose ones (of which there are many) including:

• Raphaël
• EaselJS
• D3.js
• jQuery UI
• Processing.js
• Paper.js
• Snap.svg

I also looked at diagram-oriented libraries, but they felt like too much work compared to a straightforward canvas wrapper.

• jsPlumb Toolkit
• WireIt
• JointJS - the HTML 5 JavaScript diagramming library.
• Draw2D touch: Blog
• flowchart.js
• js-graph.it homepage
• Cytoscape.js

In the end I liked Fabric best for its inbuilt support of handles and grouping, and its solid level of activity. And it worked pretty well, modulo a number of gotchas. It's a really nice piece of work, as are many of these. The JavaScript graphics library scene is definitely alive and well.

Code tour

While I don't have my client's permission to share the code, I'll sketch out the implementation for the curious.

Routes

At the top level of this standard Flask MVC app is app/routes.py, which has four URIs/controllers:

• /: app/templates/index.html template. Lists the SVG files in the fake DAO.
• /docs/<fileName>: app/templates/edit.html template. Described below.
• /docs/<fileName>/annotations: REST API endpoint for getting a document's annotations as a JSON array. Finds and reads the JSON file corresponding to the SVG file and returns it as 'application/json'.
• /docs/<fileName>/text: placeholder endpoint for calculating the text bounded by a rectangle. Pulls the rectangle bounds from the query parameters and returns a fake string.

edit.html

This view has a simple vanilla HTML editing section at the top with three hard-coded annotation types, and the editing area underneath. The latter layers the Fabric canvas on top of the SVG file, whose element is inserted dynamically via loadSvg() in app/static/edit-document.js (see below). Finally, edit.html sets the variables needed by edit-document.js, based on the Flask-injected fileName variable, and loads the Fabric and edit-document.js.

edit-document.js

This file does the heavy lifting. On load it links up the button and shortcut actions to their methods and then calls loadSvg() to load the SVG document and its annotations. It does so using an AJAX call because of a browser SVG loading bug where the SVG size is reported incorrectly until it's completely loaded. Once it's loaded, loadSvg() inserts the SVG element behind the Fabric canvas, sets its size, and then initializes the Fabric canvas, saving it as an application property on the canvas element for easy access. Once that's all done, the function calls loadAnnotations(), which performs the AJAX call to /docs/<fileName>/annotations and then creates Fabric Rect objects for each, using the correct color for the annotation type.

One tricky bit was handling the line connecting linked Rects. Lines are managed explicitly (there is no built in 'connector' feature in Fabric), so their endpoints must be dynamically adjust during rectangle moves and resizes, and pointers to/from them must be saved as properties on the Fabric objects (i.e., a Rect needs to know all of its Lines, and a Line needs to know its two endpoints' Rects).

The rest of the code manages all the fiddly aspects of the app - resizing, duplicating, linking, etc. And of course button state must be updated based on selection changes. The only other mildly interesting thing is the AJAX call to get an annotation rect's text - a straightforward call to the /docs/<fileName>/text endpoint.

User Documentation

The UI is a straightforward direct manipulation one where users work with rectangle objects. Click to select, drag to move, drag resize handles to resize, click the delete rectangle button to remove, etc. The only feature that's non- obvious is how to add and remove links between rectangles. To add a link, select exactly two rectangles with the same label and no existing link and then click the add link button. To remove a link, select two rectangles with an existing link and the click the remove link button. Keystroke shortcuts are supported for power users:

Keystroke shortcuts

Types

          1: set current type to Title
 shift  + 1: filter all but Title
          2: set current type to Abstract
 shift  + 2: filter all but Abstract
          3: set current type to Author
 shift  + 3: filter all but Author
     Escape: reset filter

Create

+: create new annotation using current type

Move & Resize

                     arrow key: move selection 1px
            shift  + arrow key: move selection 10px
           option  + arrow key: resize selection 1px
 shift  +  option  + arrow key: resize selection 10px

Delete/Duplicate

backspace|delete: delete selection
    control  + d: duplicate selection

Select

          tab: select next
 shift  + tab: select previous

Text Feedback

x: display text for selection

Proposal: An online skeptical toolbox

For some time I've been collecting project ideas to help my overall goal of "computing in the service of humanity" (a phrase I recently picked up). While I definitely want to find consulting work in this area, starting up a personal project in the meantime is important to me. Out of a bunch of ideas I've decided to start with an online skeptical toolbox.

What finally kicked me in the butt were two things: 1) The publishing of Google's paper Knowledge-Based Trust: Estimating the Trustworthiness of Web Sources (which got a good bit of coverage), and 2) the recent Committee for Skeptical Inquiry article Online Tools for Skeptical Fact Checking by Tamar Wilner. To that end I've created a GitHub project that at the moment has two files: the proposal itself and a somewhat-organized XMind mind map file skeptical-toolbox.xmind that lists some detail.

As I said in the Implementation section, the tools cover a range of complexity, which affords our quickly rolling out something useful using the simpler ones, and progressively introduce more sophisticated tools as the toolbox develops. I plan on using Python + Flask to write it in (same as I used for PeepWeather), but I'm quite open to other languages and frameworks, such as Ruby on Rails if someone steps up and convinces me. One thing in Python's favor is the popular Natural Language Toolkit with its information extraction tools, especially around entity recogition.

What do you think? It is very early days (day zero, I suppose) but I hope that sharing this will generate some thoughts. If you're interested in helping build this - awesome! Just comment here or send me a line.

(Image: Memory Belt)

Available! Versatile Software Engineer, 20+ Years Experience, Object-Oriented & Extreme Programming

After taking a break from computing in 2009 and creating a new kind of social platform - one based on treating life as an experiment - I happily returned to CS research in 2011. I was asked to re-join the AI group at UMass where I previously helped build Proximity, a platform for machine learning research. However, that position's funding has ended (as is the way of research's ebb and flow) and I'm excited to move on to the next adventure.

In those four years I got a lot done, including exploring ways to scale up the Ph.D. students' algorithms [1], coaching students in the Extreme Programming agile software development methodology that I love so much, implementing lab infrastructure improvements (I consolidated multiple aging servers into a single modern cluster, and got the lab using the commercial wiki, Confluence), and of course collaboratively designing and coding the lab's new causal learning in Python (a language I love, BTW, though I still enjoy Java).

As a pleasant surprise, before the funding ended in May, at the last minute another CS professor asked me to help his lab get some urgent work done during the summer, which I was happy to do. While there I benchmarked and started an optimization effort for their new information extraction pipeline, led the team in meeting a crucial grant deadline, and prototyped a web-based PDF annotation tool [2] for creating a gold standard to evaluate their algorithms' performance against. I was introduced to Node.js, continued learning Scala (I wrote some for my GraphX work), and JavaScript for the web app.

Looking forward, I've started my job search for my next project, with the goal of finding work that's challenging, engaging, and meaningful. I've not had to do one in a while, so the process itself is a kind of experiment. What approaches will work? What is out there? Who is doing cool work? Exciting, and a little scary.

So: if you know someone looking for a versatile software engineer with lots of experience who writes excellent code, then please - drop me a line. My LinkedIn profile is linkedin.com/in/matthewcornell.

• [1] I explored multiple MPP approaches including single-node SQL using PostgreSQL, distributed SQL using Impala on the Hadoop ecosystem and Vertica, and graph databases such as GraphX, Neo4J, and Giraph.
• [2] The prototype allows users to load a PDF file as a browser-renderable SVG, draw and edit rectangular regions overlaid on the text, and save and load them from a server. I used the standard web technologies of JavaScript and JSON, all running on Play Framework and its Java API. I used Fabric.js for the direct manipulation UI, which I really enjoyed working with (check out the awesome demos).

(Image from Fountain on Boston Common.)