Bingxi Li

Focal stacks and repeated-exposure sequences both combine many frames into one image, but for different reasons: focus stacks need frames that cover different depths, while temporal stacks need frames whose noise averages out. UIBS (unified information-budgeted selection) picks a fixed number of frames from a stack that has already been acquired, using one greedy objective with domain-specific proxies: a focus-response score for focal stacks and a shrinkage-regularized residual-covariance proxy for repeated exposures.

I evaluated it on seven insect focal stacks (519 images; some are different views of the same specimen) and three 64-frame Nikon Z8 RAW runs with matched dark frames and a separate low-noise reference, against random, uniform and single-operator greedy baselines. The baselines carry the main finding. Most of the spatial gain comes from greedy selection itself; my dual-operator score showed no consistent advantage over selecting with a single focus measure, and it can lower retention in the worst-served regions. Retention is scored with the same focus responses the selectors use, so it measures how much focus evidence survives, not the quality of a re-fused image.

An all-in-focus composite is good documentation but it erases provenance: you can no longer tell which slice supports a region, or whether that assignment is ambiguous. FocusStack2XR keeps that evidence. It computes a focus-index map, a confidence margin, a conservative validity mask and a relative focus-depth layer, and a browser viewer lets you switch overlays, browse slices and click a region to see its predicted source slice. Relative depth is treated as an inspection cue, not calibrated 3D geometry.

The “XR” is the goal, not the current system: the prototype is a desktop browser viewer and hasn’t been tried in a headset or with users. The paper’s argument is that a plausible 2.5D view can imply geometry the data don’t support, so the evidence layers have to come first.

The viewer at the top of this page is a simplified version of the same pipeline. The paper evaluates seven stacks across four specimen categories and reports where the method breaks: transparent wings and cluttered scenes.

I apply and analyze the lab’s neural-field-based reconstruction pipeline for two-photon fluorescence microscopy. A learned representation of the specimen is combined with a physical forward model of the microscope, so one optimization estimates both the specimen and the system’s aberrations, written as Zernike modes, and its point-spread function.

One problem caught my attention: a late-stage drift in the estimated optical parameters, which continued after the reconstruction loss had largely settled. I ran controlled tests of learning rates, regularization, shear handling and SLM/pupil geometry, comparing recovered Zernike coefficients with simulation references. Optimizer dynamics, model mismatch or a direction the data don’t constrain could each produce such a drift, and telling them apart is still open. I also extend an existing Gaussian-mixture-based SNR analysis across excitation powers and frame-averaging levels, testing how segmentation and masking choices change the result.

My first independent measurement project. I built a RAW-domain workflow to test a variable neutral-density filter against the ideal stop law: Nikon Z8 14-bit RAW captures converted to linear TIFFs, a homography-rectified ColorChecker, two illuminants (about 3000 K and 6000 K) and stops ND0 to ND5 with three repeats each. Transmission fell more slowly than the ideal 2⁻ⁿ (fitted log₂ slope −0.74 against an ideal −1), and colors drifted warmer with density (R/G up, B/G down), which I quantified with RGB channel ratios and CIE Lab ΔE*_ab.

It taught me to keep a pipeline linear until the last step.