The Photophile HDR Image Browser

Slide01.jpg
Subject Photophile
Keywords presentation

Slide02.jpg
Motivation
Subject Photophile
Keywords presentation
Comment Catalogers take over file management from user, requiring huge disk allocations and betraying the user's trust.
Comment Browsers can't keep track of image changes, except by changing image files
Comment Cataloging software should be snappy -- most of it is dreadfully slow.
Comment Why can't I look at image thumbnails if the files aren't on my system?
Comment I should be able to move my files around and the cataloger should find them again.

Slide03.jpg
Goals
Subject Photophile
Keywords presentation
Comment Tracking image files rather than managing them is an important philosophical and practical choice. The user never changes her data without knowing by performing an explicit "save-as" operation. The browser supports cropping, rotation, red-eye removal, and other alterations all without altering the original image file. Not only does this allow the operations to be "undone" at any future point (because they were never "done" in the first place), but Exif information and is maintained and the user does not suffer generational JPEG compression losses.

Slide04.jpg
Realized Features
Subject Photophile
Keywords presentation
Comment Browsing catalogs is very fast (over 1000 images/sec). Directories take longer (10 images/sec) due to hashing to recognize files that are already in our catalog or cached by the thumbnail manager, but this avoids the need to regenerate thumbnails, so it actually saves time in most cases.
Comment Web pages are very simplistic at this point, and photo albums are waiting for printing to work.

Slide05.jpg
Unrealized Features
Subject Photophile
Keywords presentation
Comment User-defined catalog fields are built into database, but not yet exposed on user interface.
Comment Rendering currently assumes an sRGB display. Other color spaces should be relatively simple to add, but ICC profiles don't handle HDR images.
Comment Printing is difficult due to the user-interface requirements.
Comment Linux and Windows versions will have to wait for fully functional Mac OS X version.

Slide06.jpg
Browser Layout
Subject Photophile
Keywords presentation
Comment Adding images is as simple as opening the folder, selecting the images and hitting the "Add" button. Removal is just as easy, and cut-and-paste works between catalogs as well as text editors and spreadsheets.
Comment Catalogs with 10,000+ images are no problem -- no limitations on catalog sizes or number of thumbnails.
Comment Intelligent caching keeps memory footprint small with good interactivity.

Slide07.jpg
Viewer Layout
Subject Photophile
Keywords presentation
Comment Currently supports only one viewer window, but this is likely to change in the future.
Comment Multiple images are handled with selector switch, which facilitates flip-style comparisons and sequence viewing using the slide show feature.

Slide08.jpg
Info Window Layout
Subject Photophile
Keywords presentation
Comment The info window is a slave to the browser and viewer windows, displaying information for whatever image(s) are selected or on display.

Slide09.jpg
Browser Files
Subject Photophile
Keywords presentation
Comment Most modern applications are a file management disaster. We tried to keep our file interactions as simple as possible. The only complexity not shown in this diagram is the nature of the thumbnail cache, which is actually a directory of JPEG contact sheets, with 6 320x320 thumbnails per sheet, and a database that indexes the directory's contents and tracks file use for LRU removal. This storage format allows us to keep larger thumbnails at the same cost of smaller, individual files. The thumbnail cache may be shared between users, though it doesn't support concurrent access at this time (and may never). At any time, part or all of the thumbnail cache may be blown away by the user with no harmful effect on the browser.
Comment As explained before, images are considered the precious property of the user, and can be moved around the file system or written out to CD-ROM or wherever the user wishes to take them. Since laptop computers are the ideal platform for digital images because they can travel with the camera, disk space is a premium and we cannot afford to keep extra copies of files or take up too much space with our thumbnails. The browser asks for the help the first time it loses track of a file, then finds similarly moved files with a migration path list it maintains with the user preferences.

Slide10.jpg
Browser Architecture
Subject Photophile
Keywords presentation
Comment The graphical user interface exposes the functions of a large, system-independent library that performs database management, thumbnail cache management, and image processing tasks. The cache manager underlies almost every class in the library, and assures that memory is managed in an efficient and effective way to maximize system response time. Among the imaging library's unique features is the ability to construct high dynamic range images from multiple hand-held exposures using an automatic image alignment algorithm. Fast tone-mapping and floating-point color management is supported as well.
Comment Most of the library is written in C++, though there is some legacy C code for reading and writing certain file formats via the plug-in interface, as well as some of the tone-mapping and image processing code where speed is critical.
Comment Writing an application was a whole new experience for me, and it was rather frightening in some ways to spend 4 straight months coding before getting the first menu to appear.

Slide11.jpg
Subject Photophile
Keywords presentation
Comment Images without exposure information must be combined using command-line tool at present.
Comment Paul Debevec, Jitendra Malik, "ecovering High Dynamic Range Radiance Maps from Photographs," Computer Graphics (Proceedings of SIGGRAPH 97), ACM, 1997.
Comment T. Mitsunaga and S. K. Nayar, "Radiometric Self Calibration," Proceedings of IEEE Conference on Computer Vision and Pattern Recognition, Fort Collins, June, 1999.

Slide12.jpg
LDR Exposure Registration
Subject Photophile
Keywords presentation

Slide13.jpg
Image Pyramid Alignment
Subject Photophile
Keywords presentation
Comment Details may be found in the paper, "Fast, Robust Image Registration For Compositing High Dynamic Range Photographs from Handheld Exposures," submitted to the Journal of Graphics Tools.

Slide14.jpg
Alignment Results
Subject Photophile
Keywords presentation
Comment Only horizontal and vertical shifts are computed -- rotation and warping/perspective changes are not considered. Technique is successful on about 85% of the hand-held sequences we have taken.

Slide15.jpg
Camera Response Recovery
Subject Photophile
Keywords presentation
Comment Mitsunaga & Nayar method fits samples to a low-order polynomial using standard minimization techniques. This permits compact storage and fast computation as well as fine-tuning for the exact exposure. This helps for cameras that don't always know or report precise aperture and exposure time, which includes just about all of them.

Slide16.jpg
Auto-bracket Exposures
Subject Photophile
Keywords presentation
Comment Ideally, we would take more exposures than this to get a wider range of values, but this is what the camera provides in its auto-bracket mode, which is the most convenient for hand-held photographs.

Slide17.jpg
Combined HDR Image
Subject Photophile
Keywords presentation
Comment A false color rendition of the previous 5-exposure bracketed sequence, showing the large dynamic range in the results.

Slide18.jpg
Tone-mapped Display
Subject Photophile
Keywords presentation
Comment Using our histogram adjustment technique [Larson et al 97], we are able to get all of this range into a reasonable display image.

Slide19.jpg
Best Single Exposure
Subject Photophile
Keywords presentation
Comment The middle exposure from the sequence, which loses detail in the shadows and clouds. Many outdoor images suffer from lost information in this way, since the typical camera captures 2 orders of magnitude at best (100:1), when 4 is closer to what humans are capable of perceiving (10000:1).

Slide20.jpg
Subject Photophile
Keywords presentation

Slide21.jpg
The Future
Subject Photophile
Keywords presentation

Page created Sept 17 2002 9:13:18a