Xianjie Liu

HiDe: Rethinking The Zoom-IN Method in High Resolution MLLMs via Hierarchical Decoupling

Xianjie Liu, Yiman Hu, Yixiong Zou, et al.

We discover the real bottleneck of MLLMs on high-resolution images is background interference rather than object size. HiDe uses Token-level Attention Decoupling and Layout Preservation Decoupling to achieve SOTA on V*Bench (92.1) with 75% memory reduction.

E-VAds: An E-commerce Short Videos Understanding Benchmark for MLLMs

Xianjie Liu, Yiman Hu, Liang Wu, et al.

The first e-commerce short video understanding benchmark with 3,961 videos and 19,785 QA pairs. We propose MG-GRPO multi-granularity reward strategy. E-VAds-R1 achieves +109.2% performance leap in commercial intent reasoning.

High-Precision Dichotomous Image Segmentation via Depth Integrity-Prior and Fine-Grained Patch Strategy

Xianjie Liu, Keren Fu, Qijun Zhao

Introduces monocular pseudo-depth maps as structural priors for DIS. PDFNet achieves SOTA on DIS-5K with less than 50% parameters and ~5x faster inference compared to diffusion models.

Promoting Segment Anything Model towards Highly Accurate Dichotomous Image Segmentation

Xianjie Liu, Keren Fu, Yao Jiang, Qijun Zhao

A two-stage framework integrating SAM with IS-Net for high-accuracy DIS. Advances SAM, HQ-SAM, and Pi-SAM by ~8.5%, ~6.9%, and ~3.7% Max F-measure on DIS-5K.

Uncertainty-informed Mutual Learning for Joint Medical Image Classification and Segmentation

Kai Ren, Ke Zou, Xianjie Liu, etc.

A novel UML framework for reliable medical image analysis. Uses evidential deep learning for confidence estimation and proposes Uncertainty Navigator Decoder and Uncertainty Instructor for joint classification and segmentation.