hdrgen

Create a high dynamic-range image from multiple exposures of a static scene. The input files may be JPEG or TIFF, but must be 24-bit RGB (trichromatic) images. The output is your choice of a Radiance HDR picture or a 32-bit LogLuv TIFF image. The syntax of the hdrgen command is:

hdrgen –o out_file [–r cam.rsp] [–m cachesiz] [–a] [–e] [–s stonits1] image1 [–s stonits2] image2 …

As many exposures may be given as necessary, and should ideally be spaced within two f–stops of each other. The brightest exposure should have no black pixels, and the darkest exposure should have no white pixels, but there is little point in extending beyond these limits, which may cause problems in determining the camera response function. The order of options and input files is unimportant, with the exception of the –s option, which must preceed the corresponding exposure. Following is an explanation of the options and their meanings:

–o out_file
Write high dynamic-range image to the given file. If the file has a ‘.tif’ suffix, it will be written out as a LogLuv TIFF image. If it has a ‘.exr’ suffix, it will be written out as an ILM OpenEXR image. If it has any other suffix or none at all, it will be written out as an RLE RGBE Radiance picture.

–r cam.rsp
Use the given file for the camera’s response curves. If this file exists, it must contain the coefficients of three polynomials, one for each color primary. If the file does not exist, hdrgen will use its principal algorithm to derive these coefficients and write them out to this file for later use. If a scene contains no low frequency content or gradations of intensity, it may be impossible to derive the response curve from the exposure sequence. Thus it is better to create this information once for a given camera and reuse it for other sequences.

–m cachesiz
Specify the cache size to use in megabytes. No more than this much memory will be allocated to hold image data during processing. The default value is 100. Using a smaller value may require longer processing if many input images are used, since some will need to be read in twice rather than once, but specifying a larger value than there is memory available will definitely be worse, due to virtual memory swapping.

–a Toggle automatic exposure alignment. The default value is “on,” so giving this option one time switches it off. The alignment algorithm examines neighboring exposures and finds the pixel offset in x and y that minimizes the difference in the two images. It may be necessary to switch this option off when dealing with very dark or very bright exposures taken in a tripod-stabilized sequence.

–e Toggle exposure adjustment. Normally “on,” exposure adjustment fine-tunes the scale difference between adjacent images to account for slight inaccuracies in the aperture or speed settings of the camera.

–f Toggle lens flare removal. Normally,“off,” this option is designed to reduce the scattered light from a camera’s lens and aperture, which results in a slightly fogged appearance in high dynamic-range images.

-g Toggle ghost removal. Normally “off,” this option attempts to identify and single out moving portions of the image, avoiding ghost-like or blurry results in the final image, at the expense of some added noise.

–s stonits
Set the sample-to-nits (cd/m²) conversion factor for the following image to the floating-point value stonits. This is normally determined automatically by the program from camera information stored in the Exif image header. If the image did not come directly from a digital camera, then it will be necessary to use this option for each image. If the absolute conversion is unknown, then simply pick a value for the brightest image, and increase it subsequently for each exposure in the sequence. One f-stop requires doubling this conversion factor, and two f-stops requires quadrupling.

Diagnostics

The primary failure mode for this algorithm is the one mentioned in the description of the –r option, when the exposures contain too little information to solve for the camera response function. The best solution to this problem is to take off the exposures that are very light and very dark, or to use a different sequence of images to generate a response file. This file may then be used to combine the entire set of images, since the program no longer needs to solve for the responses.

Most of the other diagnostics you will encounter are “warnings,” which means that the final image will be written, but may have problems. In particular, when the alignment algorithm fails on a hand-held sequence, some ghosting may be visible on high contrast edges in the output. Using the –a option to turn off automatic alignment will eliminate the warning, but unless the sequence was taken on a very stable tripod, the results will usually be worse rather than better.

Example

To combine all JPEG images matching a given wildcard and put into a LogLuv TIFF:

hdrgen P13351?.JPG –o testimg.tif

Author

This software was written by Greg Ward of Anyhere Software. Send comments or questions to [email protected].

References

Tomoo Mitsunaga and Shree Nayar, “Radiometric Self-Calibration,” Proceedings of IEEE Conference on Computer Vision and Pattern Recognition, June, 1999.

Greg Ward, “LogLuv encoding for full-gamut, high-dynamic range images ,” Journal of Graphics Tools, 3(1):15-31 1998.

Greg Ward, High Dynamic Range Images, web page.

Paul Debevec, web page.