Jiwon's Alcove

The DT pipeline where data sources are combined with RatioColl or EpsilonGreedy to form a balanced unified dataset.

Data Distribution Tailoring

2023
With Bohan Cui, Fatemeh Nargesian.

We study the data distribution tailoring (DT) problem, which aims to collect a unified dataset from multiple data sources such that groups of interest are adequately represented. In doing so, we ensure a baseline level of fairness in downstream tasks. We demonstrate the suboptimality of the prior paper's CoupColl algorithm and present a strictly better RatioColl algorithm. Furthermore, we generalize the prior authors' UCB algorithm for situations in which demographic statistics are not available through our EpsilonGreedy algorithm.

Red and blue points on a plane, with some red points and some blue points having a circle around them.

Fair $k$-Cover Coresets

2022
With Pranay Mundra, Yurong Yu, Fatemeh Nargesian.

We study the fair $k$-cover problem, which aims to efficiently obtain coresets such that 1) every point in the full dataset is covered by a point in the coreset at least $k$ times, and 2) points in the coreset adequately represent groups of interest. We solve the problem through a reduction to submodular optimization. Our $\mathrm{F}\mathcal{K}\mathrm{C}$ coresets reduce accuracy disparity while generating coresets at a fraction of the cost as SOTA.

The pipeline. An SDR image is made brighter and darker, then denoising is applied. The resultant three images are combined with Mertens' fusion.

Single Exposure Fusion

2022
With Yuhao Zhu.

code  report 

Photographers can extend the perceptual dynamic range of a photograph by utilizing the exposure fusion method, in which multiple photographs taken at different exposures are combined. However, this technique is susceptible to blur if the camera or subject moves.

I developed an alternate SDR to HDR pipeline which only requires one medium-exposure SDR photograph and will not be blurred due to camera or subject's movement. Unlike existing alternatives, single exposure fusion is a simple combination of classic denoising and exposure fusion algorithms that does not require machine learning.