Golan Levin and Collaborators

Flong Blog + News

Open-Source Panoramic Video: Bloggie + OpenFrameworks & Processing

23 June 2010 / code, project

I’m pleased to announce the release of free code and tools, created in openFrameworks and Processing, for unwarping the panoramic videos produced by the popular Sony Bloggie solid-state camera. The inexpensive Bloggie ($160-190), in many ways similar to the Flip, is a perfectly decent little pocket camera with one exceptionally notable feature: it allows for the creation of 360-degree panoramic digital video, at relatively high resolution, at a consumer price-point. Here’s an example:

Obviously, this “donut vision” (annular) video needs to be transformed into something more useful, such as a long panoramic strip — a process variously called unwarping, dewarping, inverse warping, remapping, reprojection, rectilinearization, or simply distortion correction. Although Sony released unwarping software for Microsoft Windows, they — inexplicably — didn’t release Bloggie unwarping software for Mac OSX. Well, that’s problematic. Perhaps because panoramic imaging is considered a “niche” market, commercial tools for panoramic unwarping can be quite pricey. And although Christophe Maillot has lately released ThreeSixZero, a free OSX tool for viewing Bloggie panoramas, his project is not open-source and doesn’t export converted panoramic video! That’s the motivation for the releases presented here. Fortunately, the mathematics for polar-to-cartesian geometry conversion are relatively straightforward. Behold, a 360-degree panoramic video, remapped to a rectilinear perspective with our BloggieUnwarper software:

Projects and Source Code

Below are links to BloggieUnwarper, my open-source code projects for unwarping the panoramic videos produced by the Sony Bloggie. These projects are created in openFrameworks (a C++ library for creative coding) and Processing (a Java-based programming environment for people who want to create images, animations, and interactions). Both OpenFrameworks and Processing are free, open-source, cross-platform toolkits, optimized for visual computing, which take most of the headache out of creating graphically-oriented software. Both programming environments work across multiple operating systems (Mac, Windows, and Linux), and multiple compilers (e.g., openFrameworks compiles in XCode, Visual Studio, Code::Blocks, and more). If you’re interested in interactive art or computational design, these environments are a great way to start making projects fast.

Here’s a screenshot of the openFrameworks BloggieUnwarper software, showing the original (annular) video at upper left; the reprojected panoramic strip at the bottom; and (in the upper right) a view of the panorama from within a texture-mapped 3D cylinder:

Complete source code can be found here:

  • BloggieStillUnwarp.zip (for Processing; 1.3Mb download) — This is a single-page Processing sketch which performs inverse warping on a single panoramic still image stored in the sketch’s data folder. The program supports both nearest-neighbor and bicubic interpolation, and can save out the unwarped panorama to disk. To get started, download the Processing environment here; unzip this project into your “Sketchbook” folder (in “Documents”); open the sketch in the Processing development environment, and press play.
  • BloggieVideoUnwarp.zip (for Processing; 20.2Mb download) –This is a single-page Processing sketch which unwarps a Quicktime video stored in the sketch’s data folder, and exports an unwarped version of the video — either as a new Quicktime movie, or as a folder full of still images. Parameters for the unwarping (such as the optic center) are loaded from a settings file. The audio from the original file is (regrettably) not maintained; see my notes below about audio. (Please note that the bulk of this download is a 20Mb demonstration video.)
  • BloggieUnwarpOF.zip (for OSX & openFrameworks, 52.7Mb download) — This is a complete openFrameworks project for viewing, projecting, dewarping, and exporting panoramic videos shot with a catadioptric reflector like the Sony Bloggie’s. This project is built for Mac OSX 10.6.3, using XCode, and openFrameworks v.0.061, which is available here; be sure to download the FAT version. The project makes use of the ofxOpenCv and ofxXmlSettings addons which are included in the download. To avoid compilation errors, install the project at the same hierarchical directory level as the various “example” apps which come with the FAT download. There is some additional information about getting OF set up on Snow Leopard here. This project is available two ways:

More About the BloggieUnwarper

Here’s some more information about panoramic video on the Sony Bloggie, our app, and its implementation.

Sourcing the Bloggie and its Panoramic Lens Attachment.
The Sony Bloggie MHS-PM5K (about $160-180) comes with an included 360-degree panoramic lens attachment. You may hear this accessory erroneously referred to as a “fisheye lens“, but it’s actually a hyperbolic or parabolic mirror (with an optical path folded by a flat mirror, as it turns out) known as a catadioptric reflector. This panoramic lens attachment, which easily clips on and off the body of the camera, is the black piece shown above. The attachment is Sony Part Number X25467051, a “360 VIDEO ACCESSORY”, labeled VCL-BPP1 on its body. The attachment really does come with the camera! But as an isolated component, this accessory is also available here through Sony’s services site (Tel. 1-800-488-7669) for $82.95 USD. Now given that this is, like, a $5 piece of plastic, this is still a minor bummer, but oh well: competing mirror-based optics for one-shot 360-degree imaging, like the Palnon parabolic C-mount panoramic lenses from JBC, the Go Pano from EyeSee360, the OmniAlert360 from RemoteReality, the PentaOne P360 series ($~1000), the EggSolution EGG Lens (~$2000), or the 0-360 Panoramic Optic ($495), can cost hundreds or even thousands of dollars; there’s simply nothing else quite like the Bloggie at the low end. In any case, it’s worth pointing out that you can duct-tape this Sony pano accessory onto just about any micro-format camera if your application demands a different hardware scenario. I made relatively decent panoramic images by simply holding it in front of the cameras on my laptop and my mobile phone, and I also had good luck affixing the mirror module onto standard 12mm “micro-video” lenses such as those used in many web-cams. If you’re interested in doing something low-budget like DIY-Streetview, the Bloggie and/or this lens attachment is the camera for you.

Limits on the Resolution of Resulting Panoramas.
The Sony Bloggie is capable of recording non-panoramic HD video at resolutions up to 1920×1080 (30fps). Unfortunately (and this isn’t well-advertised), the Bloggie only records panoramic digital video at 1280×720. In video mode, the camera auto-detects the panoramic lens attachment (magnetically, it turns out) and switches to the lower resolution. Moreover, by the time the unwarping process has removed the wasted black pixels (inside and around the warped video annulus), it becomes clear that the maximum effective vertical resolution of the Bloggie’s panoramic video is only about 256 pixels, which may or may not suit your purposes. Used in still-picture mode, the Bloggie is able to create panoramic source images with its full (5 megapixel, 2592×1944) resolution; the effective vertical resolution of its panoramic stills is then 512 pixels. Since the Bloggie lens accessory has a vertical FOV of approximately 54.5 degrees (according to our measurements; others have measured the vertical field of view at 52.5°), panoramas produced with these optics should have aspect ratios close to (360/54.5) ≈ 6.6:1. For panoramas produced with the BloggieUnwarper, I therefore recommend dimensions like 1280×196, 1680×256, or (for 5MP stills) 3360×512. (Note that the Bloggie has a 1/2.5” CMOS imaging sensor and F3.6 optics. This means: expect relatively poor low-light performance, and some very exotic rolling-shutter artifacts on fast-moving subjects after your source video is unwarped.) Update: in a new blog post, “Modding the Bloggie Panoramic Lens Accessory“, I detail a simple modification to the Bloggie accessory which enables 1920×1080 panoramic video recording.

An Opportunity for Real-Time Panoramic Video
The BloggieUnwarper software can compute unwarped panoramas in real-time or faster. And happily, the Bloggie camera can also produce real-time panoramic video, through its analog NTSC/PAL composite video output, using the proprietary Sony HandyCam Audio/Video Cable (VMC-15FS A/V) which comes with the camera. This interlaced analog video signal produced by the Bloggie is then suitable for 720×480 capture via standard analog capture cards such as the Dazzle. An opportunity thus exists for (e.g.) interactive computer-vision based applications that use 360-degree sensing. To be perfectly clear, the Bloggie does not support use as a real-time USB webcam; you’ll need to run its analog video output through a capture card (or fit the Bloggie lens attachment onto some other camera) if you intend to build apps that incorporate real-time panoramic video.

The code projects provided here are configured for unwarping Bloggie videos stored on disk — but modifying these apps to use real-time video is straightforward; you’ll simply need to swap out the source of input pixels from the pre-stored video to the live video input:

  • If you’re working in openFrameworks, integrate the ofVideoGrabber class into the project, as demonstrated by the movieGrabberExample example included in the FAT download, to understand how to access the live video stream and its pixels.
  • For Processing, on the other hand, integrate the Capture class within the Video library, taking a careful look at the included GettingStartedCapture and Mirror examples to understand how to access the live video and its pixels.

Special Considerations, Features, and Known Bugs
Unwarping panoramic video from the Bloggie has a few special considerations, mentioned here.

  • Handling Inexact Centering of the Panoramic Optics.
    Sad to say, the Bloggie’s removable panoramic attachment is a wee jiggly. As a result, it’s unwise to assume that the optical center of the video donut is located precisely at the location [width/2, height/2]; indeed, the optical center has a movement range of approximately ±2% of the image dimensions. This has significant implications for an unwarping tool; even small differences of a quarter pixel can cause an undesirable sinusoidal “wobble” that runs the length of the panorama. The apps provided here allow for the user to adjust, set and save the center point of the video annulus, with sub-pixel accuracy.
  • Handling Corrections for Radial/Vertical Nonlinearity
    If the Bloggie had a ‘perfect’ catadioptric mirror, there would be no vertical distortion in the unwarped panoramas. Unfortunately, our careful measurements of rectilinear grids photographed with the Bloggie’s panoramic lens attachment proved otherwise; we discovered a minor nonlinear distortion in the radial/vertical dimension of the Bloggie lens. We also discovered that the Bloggie panoramic lens has an asymmetrical vertical field of view, with approximately 80% of the image falling above the horizon, and 20% below. Taking both perspectival distortion and the Bloggie’s off-center vertical optics into account, we computed this distortion, modeled it with a 2nd-order polynomial, and have corrected for this in our unwarping procedure. You’ll need to recalculate these fitting parameters if you wish to use this software with a different catadioptric mirror.
  • Providing Azimuthal/Horizontal Shifting
    When unwarping an annular video to a horizontal panoramic strip, the placement of the “zero angle” becomes an important artistic choice. Modifying this angle has the effect of determining the horizontal placement of objects in the scene. BloggieUnwarper allows for user-determined shifting of this kind.
  • Bugs with Preserving Source Audio
    As of this writing, BloggieUnwarper doesn’t handle audio nearly as well as I would like. The Processing version does not preserve the source audio at all; if you want to keep it, you’ll need to use a tool like FinalCut or AfterEffects to recombine the original source audio with the corrected panorama video. The openFrameworks version is a little more strange. It currently produces “non-standard” videos, which contain the original audio alongside the unwarped video, and are playable in Quicktime7+ on the Mac, but which fail to play in Windows, and are rejected by Vimeo for unspecified incompatibilities. I have successfully circumvented this by re-compressing the results using Quicktime and/or MPEG Streamclip.

A brief note about speed and library dependencies.
To accomplish fast (better than real-time) inverse warping, we compared four methods: a well-optimized, hand-coded inverse lookup table; the cvRemap() function available within the free OpenCV library; the fwiRemap() function in AMD’s free FrameWave signal processing library; and the ippiRemap() function from the commercial Intel Performance Primitives (IPP) 6.0 libraries. That’s right — we had a shootout. All of our implementations were coded for 3-channel (RGB) images, and for each method we tested both Nearest Neighbor and Bicubic interpolation methods. We set compiler optimizations to level -O3 in XCode and compiled for 10.6|Release. Under these conditions, the quick summary of the libraries’ relative performance was as follows:

  • IPP ippiRemap: 1 time unit
  • OpenCV cvRemap: 2 time units
  • Framewave fwiRemap: 4 time units
  • Hand-coded version: 2.5-10 time units, depending on the extent of optimization

In other words: IPP was the fastest; OpenCV was twice as slow as IPP; Framewave was four times as slow; and (short of getting all medieval with assembly code), my own best hand-coded algorithm was about 25% slower than OpenCV. Based on the above shootout, we elected to use OpenCV with our openFrameworks build (see the ofxOpenCV addon example, which comes with the download), and a hand-coded unwarping algorithm for the Processing version. Researchers who are serious about speed may wish to consider the (non-free, non-open-source) IPP libraries. As an aside, OpenCV is documented as being able to “take advantage of IPP if it is available“, but this feature of OpenCV is being phased out.

Acknowledgments
Ivan Poupyrev first alerted me to the existence of the Sony Bloggie. Solomon Bisker helped port my original Processing sketch to openFrameworks, and also set up the four-way shootout between IPP, Framewave, OpenCV, and my hand-coded algorithms. Amisha Gadani tracked down purchasing details for the X25467051 lens accessory. Kyle McDonald patiently walked me through learning Git, and collaborated in many stimulating discussions and experimentation related to low-cost panoramic video. This (side-)project was supported by the STUDIO for Creative Inquiry at Carnegie Mellon University, with the help of grants from the Fine Foundation and the National Endowment for the Arts. This project has been featured at Hack-a-Day.

Keywords
Golan Levin, Flong, Processing, openFrameworks, #Processing, #openFrameworks, Sony Bloggie camera, 360° video, 360-degree pano, panoramic video, catadioptric reflector, hyperbolic reflector, parabolic mirror, unwarp, dewarp, inverse warp, remap, reprojection, rectilinearization, distortion correction, free open-source code, Java, C++, software, free video tools, optics, lens, attachment, one-shot panoramic imaging, download, Bloggie software for Mac OSX, free Bloggie for MaxOS X, Mac OS X.

References
A Theory of Single-Viewpoint Catadioptric Image Formation“, S. Baker and S.K. Nayar. International Journal on Computer Vision, Vol.35, No.2, pp.175-196, Nov, 1999. [PDF]

Image Warping / Distortion“, P. Bourke. http://local.wasp.uwa.edu.au/~pbourke/miscellaneous/imagewarp/, Dec. 2002.

Licence
This software is licensed under the CC-GNU LGPL version 2.1 or later.
CC-GNU LGPL


Comments are closed.

« Prev post:
» Next post: