Research

Recently, research in 3D reconstruction shifts from achieving consistency in mere appearance and geometry to attaining physically plausible models of the scene or the object. For this problem, while test-time optimization approaches takes hours to optimize reasonble physical parameters of even a single object, the generalizability of feed-forward approaches is too limited. We are currently striving to overcome the shortcomings of both approaches and provide a simulation pipeline that is easily generalizable.

MixCuBe — a cross-cultural VQA benchmark built via a novel image-augmentation pipeline, used to evaluate the cultural bias of SOTA multimodal LLMs in mixed-cultural settings.

Bringing a transformer-based rendering pipeline (Microsoft's RenderFormer) from O(N²) to linear time complexity via Performer (FAVOR++) attention.

A novel graph-ML architecture learning temporal and spatial patterns of brain activation from fMRI images, with two downstream tasks: brain-activation link prediction during language tasks, and performance prediction from neural patterns (graph regression).