Robert, let me first say that I greatly respect the work you're doing (which Doug Ellison pointed me to) towards colorizing and calibrating the MER images.
Just to explain where I'm coming from, I'm an amateur colorizer myself, and I've been working on my own on algorithms that can take input of images calibrated to radiance (relative is ok, absolute would be best), interpolating a spectrum from the 6 L2->L7 frames and building color information from that, using the CIE XYZ colorspace as a basis of the model. As of right now, the input is completely uncalibrated and is available at:
http://www.lyle.org/~markoff/

After some early, failed, efforts at auto-calibrating the raw images to true brightness, I gave it a rest to concentrate on the colorization algos. Your recent work here shows great promise towards calibration of the images to a good standard of uniformity without additional information from NASA regarding the actual exposure times of the images in question. Though, I do have a couple questions regarding the method by which you are achieving the autocalibration.

"...uses the published ground calibration data for the filters, plus dynamic rebalancing to undo the effects of autoexposure in the cameras"

Out of curiosity, is this based on Bell el al "The Mars Exploration Rover Athena Panoramic Camera (Pancam) Investigation" and Maki et al "Mars Exploration Rover Engineering Cameras"? Both of these papers were very useful to me in my own efforts, though the 'holy grail' of calibration papers (referenced from the Bell paper) is the one I can't seem to find publicly:
Bell, J.F., III et al., Mars Exploration Rover (MER) Project Pancam and Engineering Cameras Calibration Report, JPL Document D-19826, MER-420-6-700, Jet Propulsion Laboratory, Pasadena, CA, in preparation, 2003b. I wonder if you or anyone else has seen this around anywhere?

I've been reading up on the methods of autoexposure and comparing the actual autoexposure values used for separate, 6 image (L2->L7) image sets that have already been taken using the PDS tags included with the files distributed as part of Maestro ( http://mars.telescience.org ) datasets. Please excuse any parts of this which I may have misunderstood.

The numbers below refer to the L2-L7 set taken of post-brush, pre-rat Humphry Spirit, Sol 55 L2 reference 2P131251640EFF1147P2583L2M1 which used autoexposure, and not reference to the exposure table. This is not always the case, as the exposure table can be saved from one autoexposure set and used on subsequent shot (parts of a panorama for example) which has similar content and lighting. Justin Maki (side note, how cool is it that this guy responds to email?) replied to my query about this by saying "Yes the exposure time tables are being used extensively, but the values in the tables are dynamic and are automatically updated based on scene content." Unfortunately, this means that certain scenes are not using the autoexposure formula's at all (and therefore wouldn't necessarily exhibit the same histogram characteristics), but rather a fixed exposure ratio based off an autoexposed L2 frame, which could lead to some problems if the assumption is that every L7 frame meets the autoexposure characteristics...but i digress, as I said, the numbers below *are* autoexposure figures.

The first of the values used for a purely autoexposure features is:
Auto_exposure_data_cut: provides the DN value which a specified fraction of pixels is permitted to exceed. The fraction is specified using the auto_exposure_pixel_fraction keyword.
For all frames L2->L7, this value was set to 2500 (out of 4096 DN on the 12 bit images)
Auto_exposure_pixel_fraction: provides the percentage of pixels whose value is higher than the auto_exposure_data_cut keyword.
For the L2 frame this value was set to 10, for all subsequent frames L3->L7 this was set to 20.

I wonder how your calibration can handle images whose autoexposure is set to different (and when working from the raw images, and not those with corresponding PDS tags, unknown) fractions. The difference between 10 and 20% of the pixels exceeding ~156 DN (in the frame of reference of the 256 bit images received on Earth) could account for the low spectral characteristic of the L2 frame in your spectrometer graph.

I think that most all the calibration being done at JPL, for the eventual release of "Calibrated to Reflectance" files promised to be part of the PDS data release of Pancam files, is likely being done from the other direction. They can probably take exposure time (and related functions like the temperature of the pancam at the time of imaging) to remove the flatfielding, dark current, and adjusting radiance as a function of exposure time. Thankfully, from what Bell mentioned in his Pancam paper, DN/second of exposure time is very strongly linear with Radiance once the rest of the calibration has taken place. Unfortunately, we don't have exposure times for the raw images! Hence my interest in alternative methods of calibration, such as the one you are working on.

For reference, if this would assist you in your own verification, the true exposure times are included as well in the PDS tags of the Maestro datasets

Again, post-brush, pre-rat humphery Spirit, Sol 55 L2 reference 2P131251640EFF1147P2583L2M1
L2: 225.28 ms
L3: 353.28 ms
L4: 389.12 ms
L5: 1228.8 ms
L6: 1950.72 ms
L7: 6681.6 ms

This massive difference in exposure time is more extreme than it appears though, as the responsivity to the pancam + filter is much lower for L7 frames than it is for L2, but even with this taken into account, there is still a very heavy relative overexposure of the "blue" frames.
Here are the relative responsivity of Spirit, Left Pancam + filters as listed in Bell's "The Mars Exploration Rover Athena Panoramic Camera
(Pancam) Investigation" ( http://europa.la.asu.edu:8585/PGG/greeley/courses/pdf/bell_2003.pdf ) in units of (W/m2/nm/sr)/(DN/sec):
L2 3.739E-06
L3 6.321E-06
L4 7.182E-06
L5 1.117E-05
L6 1.242E-05
L7 2.911E-05

As you can see, the L5 and L6 frames, the ones you show as having the highest value in your spectral chart, show the lowest responsivity (yielding the fewest DN per second of a given scene).

put them together, and you get a feel for the relative intensity of light at a given wavelength represented by a 1 DN step in the original 12bit image normalized to L4, which according to this formula, would have at those exposure times, the highest radiance/DN value
L2 0.8992
L3 0.9694
L4 1.0000
L5 0.4925
L6 0.3450
L7 0.2360

This very rough calculation, which accounts only for exposure times and does nothing with any of the other calibration effects (dark current, bias, flatfielding...many of which would skew the longer length frames brighter)
but it does show that a given a single pixel value step from 0-4096 in the original 12 bit L4 frame would represent twice the radiance of the comparable step in an L5 frame. This may, in part, account for your method of calibration showing L4 at half the value of L5 for the relatively bright reflective half of the brushed target.

I don't wish to post here sounding like an expert in this, because I'm not. I'm learning as I go, and acknowledge that I could be completely wrong about my understanding of everything above. Furthermore, you might already know everything I just described and already taking them into account within your method for all I know. I simply post here to bring this information into the discussion, in the hopes that an accurate calibration technique can be used for the time being, to approximate true color. I'll have a good amount of faith in the PDS calibrated files, once they come out, but in the mean time, I'd love to know what these scenes look like!

Thanks again for your efforts