Dynamic Resolution Routing for Efficient Egocentric Grounding
NeurIPS 2026
SmartRes reduces visual processing costs for egocentric grounding by combining a low-resolution global view with selectively encoded, high-resolution object-centric regions.
My work spans small object detection, visual representation learning, and efficient vision-language models, with recent work on adaptive resolution for efficient vision-language understanding.
NeurIPS 2026
SmartRes reduces visual processing costs for egocentric grounding by combining a low-resolution global view with selectively encoded, high-resolution object-centric regions.
ICCV 2025
We address unstable gradients in small-object localization by reformulating localization as classification, with uncertainty minimization and refinement to improve localization precision.
CVPR 2025
SET improves tiny object detection by suppressing high-frequency background noise and using adversarial perturbations during training to strengthen object features.
ACM TOMM
P4Q adapts post-training quantized CLIP to downstream tasks through learnable prompts, low-bit adapters for image–text feature alignment, and distillation from a full-precision teacher.
Pattern Recognition
We improve robustness to noisy annotations in tiny object localization by modeling localization errors with normalizing flows and downweighting gradients from high-uncertainty samples.
AAAI 2024
AQ-DETR improves low-bit detection transformers through auxiliary queries during training and layer-by-layer distillation to reduce attention quantization errors.
ACPR 2023
Q-YOLO addresses accuracy degradation in post-training quantized YOLO detectors by calibrating activation clipping thresholds to minimize quantization error under imbalanced activation distributions.
* Equal contribution.