CamReasoner: Reinforcing Camera Movement Understanding via Structured Spatial Reasoning [Paper] [Code]

Hang Wu, Yujun Cai$^{\dagger}$, Zehao Li, Haonan Ge, Bowen Sun, Junsong Yuan, Yiwei Wang

Arxiv 2026

• Structured Reasoning Paradigm: We propose the Observation-Thinking-Answer (O-T-A) paradigm, reformulating camera movement understanding from black-box classification into a structured ego-motion inference task that decouples camera trajectories from dynamic scenes.

• Inference Trajectory Suite: We construct a Large-scale Inference Trajectory Suite with 18k SFT and 38k RL samples, transforming static labels into dense reasoning chains to instill spatio-temporal cinematic logic into MLLMs.

• RL-driven Logical Alignment: We are the first to employ Reinforcement Learning for logical alignment in this domain, achieving state-of-the-art 78.4% and 74.5% accuracy in binary classification and visual question answering tasks.

RefineShot: Rethinking Cinematography Understanding with Foundational Skill Evaluation [Paper] [Code]

Hang Wu, Yujun Cai, Haonan Ge, Hongkai Chen, Ming-Hsuan Yang, Yiwei Wang$^{\dagger}$

Arxiv 2025

• Benchmark refinement: Enforces consistent option granularity, unified evaluation dimensions, and mutual exclusivity for greater dataset reliability.
• Baseline analysis: Thoroughly evaluates ShotVL, revealing weaknesses in reasoning, prompt adherence, and output consistency.
• Evaluation expansion: Adds a protocol assessing both task-specific performance and core model competencies, enabling more balanced and robust comparisons.

DiMo-GUI: Advancing Test-time Scaling in GUI Grounding via Modality-Aware Visual Reasoning [Paper] [Code]

Hang Wu, Hongkai Chen$^{\dagger}$, Yujun Cai, Chang Liu, Qingwen Ye, Ming-Hsuan Yang, Yiwei Wang

EMNLP 2025 Main Conference

• We propose DiMo-GUI, a training-free framework that can be seamlessly integrated as a plug-and-play component into any GUI agent.
• Without requiring additional training or external data, DiMo-GUI effectively enhances grounding performance across various GUI tasks.

Research on Perception-oriented High Dynamic Range Imaging Systems

National-level Innovation Project

• We treat artifacts in HDR images as detectable entities, explicitly detect and suppress them to enhance HDR quality.
• National-level innovation project at Tongji University, with a funding of 10,000 RMB.