Research in video understanding, object-centric reasoning, and efficient visual learning.
My name is shown in bold. Download all BibTeX · Google Scholar
arXiv preprint, 2026
SlotNarrative links recurring object observations through a lightweight, parameter-free memory, then represents each object with identity and state tokens. It provides a compact interface of 144 allocated visual-token positions for a frozen Video-LLM, independent of the number of sampled frames.
arXiv preprint, 2026
TOC-Bench evaluates whether Video-LLMs preserve object identity, state, and temporal continuity through occlusion, disappearance, reappearance, and interactions. It contains 2,323 human-verified questions over 1,951 videos, covering 10 diagnostic dimensions with object-track-grounded evidence.
Journal of Electronic Imaging, 34(3), 033047, 2025
Semi-RSTOD combines contextual feature refinement (CFR) with an adaptive perception loss (APL) for remote-sensing tiny-object detection. CFR uses multiscale cross-view features to preserve salient regions and fine details, while APL accounts for object size and class variation to address imbalance in pseudo-labels.
IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pp. 6720–6724, 2024
LCPAQ combines hardware-aware profiling, Hessian-based sensitivity analysis, and Pareto selection for mixed-precision quantization. Integer programming and a low-cost proxy search help select layer-wise quantization settings for resource-constrained deployment.
