arXiv subjects

Siwoo Lee

Publications and source records attributed to Siwoo Lee.

6 recordsLinked to original sources

arXiv2026

M2Depth: Unifying Monocular Depth Foundation Priors with Multi-View Stereo

Deep learning-based Multi-View Stereo (MVS) has advanced significantly but often generalizes poorly to unseen scenes, particularly in occluded areas or regions with limited view overlap. To mitigate this, recent approaches integrate Depth Foundation Models (DFMs) into MVS pipelines to provide monocular depth priors. However, existing methods typically rely on a static, one-way fusion scheme, which fails to fully exploit the complementary strengths of both modalities. We propose a novel framework that overcomes this limitation by tightly coupling a DFM with a cascade MVS pipeline through a bidirectional mutual refinement strategy. Our method leverages MVS depth to resolve the scale ambiguity in monocular predictions, while the monocular depth, in turn, enhances the structural completeness and fine-grained detail of the MVS estimate. Furthermore, we introduce a prior-guided cost volume refinement mechanism that effectively integrates multi-view and monocular information via attention-based fusion and discretized depth bins, thereby promoting local geometric consistency. Extensive experiments demonstrate that our method outperforms state-of-the-art MVS approaches on standard benchmarks, producing more complete and generalizable depth maps with sharp boundaries. Furthermore, although not explicitly designed for sparse-view settings, our framework generalizes remarkably well, competing favorably with even dedicated sparse-view methods while maintaining a superior accuracy-efficiency trade-off.

cs.CV↗

arXiv2026

PR-IQA: Partial-Reference Image Quality Assessment for Diffusion-Based Novel View Synthesis

Diffusion models are promising for sparse-view novel view synthesis (NVS), as they can generate pseudo-ground-truth views to aid 3D reconstruction pipelines like 3D Gaussian Splatting (3DGS). However, these synthesized images often contain photometric and geometric inconsistencies, and their direct use for supervision can impair reconstruction. To address this, we propose Partial-Reference Image Quality Assessment (PR-IQA), a framework that evaluates diffusion-generated views using reference images from different poses, eliminating the need for ground truth. PR-IQA first computes a geometrically consistent partial quality map in overlapping regions. It then performs quality completion to inpaint this partial map into a dense, full-image map. This completion is achieved via a cross-attention mechanism that incorporates reference-view context, ensuring cross-view consistency and enabling thorough quality assessment. When integrated into a diffusion-augmented 3DGS pipeline, PR-IQA restricts supervision to high-confidence regions identified by its quality maps. Experiments demonstrate that PR-IQA outperforms existing IQA methods, achieving full-reference-level accuracy without ground-truth supervision. Thus, our quality-aware 3DGS approach more effectively filters inconsistencies, producing superior 3D reconstructions and NVS results. The project page is available at https://kakaomacao.github.io/pr-iqa-project-page/.

cs.CV↗

arXiv2025

Are neural scaling laws leading quantum chemistry astray?

Neural scaling laws are driving the machine learning community toward training ever-larger foundation models across domains, assuring high accuracy and transferable representations for extrapolative tasks. We test this promise in quantum chemistry by scaling model capacity and training data from quantum chemical calculations. As a generalization task, we evaluate the resulting models' predictions of the bond dissociation energy of neutral H$_2$, the simplest possible molecule. We find that, regardless of dataset size or model capacity, models trained only on stable structures fail dramatically to even qualitatively reproduce the H$_2$ energy curve. Only when compressed and stretched geometries are explicitly included in training do the predictions roughly resemble the correct shape. Nonetheless, the largest foundation models trained on the largest and most diverse datasets containing dissociating diatomics exhibit serious failures on simple diatomic molecules. Most strikingly, they cannot reproduce the trivial repulsive energy curve of two bare protons, revealing their failure to learn the basic Coulomb's law involved in electronic structure theory. These results suggest that scaling alone is insufficient for building reliable quantum chemical models.

physics.chem-ph↗

arXiv2025

MuAPBEK: An Improved Analytical Kinetic Energy Density Functional for Quantum Chemistry

Orbital-free density functional theory (OFDFT) offers a true realization of the Hohenberg-Kohn theorems, enabling full quantum-mechanical studies of electronic systems based solely on electron densities. However, OFDFT remains limited by the difficulty of formulating accurate kinetic-energy density functionals. In this paper, we substantially enhance the accuracy of OFDFT energies and densities by tuning, during density initialization, the parameter $\mu$ of the APBEK functional, which arises in the second-order gradient expansion of the kinetic energy for semiclassical neutral atoms. We augment this parameterized APBEK functional with two physically-motivated, non-empirical corrections derived from Kato's cusp condition and the virial theorem. The resulting functional, which we call MuAPBEK, is benchmarked against Kohn-Sham density functional theory (KSDFT) on atoms, organic molecules from the QM9 dataset, and the anti-malarial drug artemisinin. MuAPBEK achieves much lower energy errors than standard APBEK and Thomas-Fermi-von-Weizsacker functionals, even when the latter two are evaluated on converged KSDFT densities. Its mean absolute energy errors on atoms and molecules are 161 and 122 kcal/mol, respectively, indicating that MuAPBEK's errors do not scale with system size. MuAPBEK also yields accurate densities, with a mean integrated absolute density error of 1.8 electrons for molecules. Importantly, one step of our density optimization scheme is at least ten times faster than a single KSDFT self-consistent field cycle and exhibits a lower-order computational time complexity of $O(N^{1.96})$ with respect to system size, $N$. Our results indicate that highly-accurate OFDFT for large-scale quantum simulations beyond the practical limits of KSDFT is within reach.

physics.chem-ph↗

arXiv2024

High-Tc superconductor candidates proposed by machine learning

We cast the relation between the chemical composition of a solid-state material and its superconducting critical temperature (Tc) as a statistical learning problem with reduced complexity. Training of query-aware similarity-based ridge regression models on experimental SuperCon data achieve average Tc prediction errors of ~5 K for unseen out-of-sample materials. Two models were trained with one excluding high pressure data in training ("ambient" model) and a second also including high pressure data ("implicit" model). Subsequent utilization of the approach to scan ~153k materials in the Materials Project enables the ranking of candidates by Tc while accounting for thermodynamic stability and small band gap. The ambient model is used to predict stable top three high-Tc candidate materials that include those with large band gaps of LiCuF4 (316 K), Ag2H12S(NO)4 (316 K), and Na2H6PtO6 (315 K). Filtering these candidates for those with small band gaps correspondingly yields LiCuF4 (316 K), Cu2P2O7 (311 K), and Cu3P2H2O9 (307 K).

cond-mat.supr-con↗

arXiv2023

Autonomous data extraction from peer reviewed literature for training machine learning models of oxidation potentials

We present an automated data-collection pipeline involving a convolutional neural network and a large language model to extract user-specified tabular data from peer-reviewed literature. The pipeline is applied to 74 reports published between 1957 and 2014 with experimentally-measured oxidation potentials for 592 organic molecules (-0.75 to 3.58 V). After data curation (solvents, reference electrodes, and missed data points), we trained multiple supervised machine learning models reaching prediction errors similar to experimental uncertainty ($\sim$0.2 V). For experimental measurements of identical molecules reported in multiple studies, we identified the most likely value based on out-of-sample machine learning predictions. Using the trained machine learning models, we then estimated oxidation potentials of $\sim$132k small organic molecules from the QM9 data set, with predicted values spanning 0.21 to 3.46 V. Analysis of the QM9 predictions in terms of plausible descriptor-property trends suggests that aliphaticity increases the oxidation potential of an organic molecule on average from $\sim$1.5 V to $\sim$2 V, while an increase in number of heavy atoms lowers it systematically. The pipeline introduced offers significant reductions in human labor otherwise required for conventional manual data collection of experimental results, and exemplifies how to accelerate scientific research through automation.

physics.chem-ph↗