Ming Dai戴铭

SimpleSearch-VL is an efficient, reliable, and practical framework for multimodal agentic deep search. It improves the agent’s search-and-verification process with Factorized Adaptive Rollout, evidence-verified reasoning, and self-summarized visits, achieving strong search behavior from only 5K SFT trajectories and 2K RL prompts.

ScanFocus is a coarse-to-fine framework for spatio-temporal video grounding, decoupling global spatio-temporal scanning from local boundary refinement. It introduces Deformable Semantic-Motion Fusion for coarse proposal generation and SGTA for dense boundary-focused temporal modeling, improving precise timestamp regression.

VideoSEG-O3 is a multi-turn RL framework for RVOS, actively exploring temporal intervals and keyframes through temporal-spatial CoT instead of relying on fixed sampled frames. It further introduces SEG-aware logit calibration and a decoupled thinking trace, aligning token-level policy optimization with pixel-level mask quality.

DeRVOS decouples trajectory generation and multimodal understanding, with TAIS aligning and selecting instance trajectories for robust RVOS.

MomentSeg is a MLLM method, which unifies temporal grounding and segmentation, enabling key-frame extraction without relying on any external models. In addition, we introduce a novel [FIND] token, which allows the model to perform temporal grounding without requiring any additional timestamp encoding.

PropVG achieves end-to-end two-stage visual grounding, overcoming the traditional drawbacks of previous two-stage approaches that relied on external detectors and were often associated with slow inference and limited performance.

DeRIS analyzes a key bottleneck in visual grounding—Cognition. It decouples the VG task into perception and cognition components, and integrates them effectively through a loopback synergy mechanism.

C3VG investigates the consistency prediction problem in REC and RIS, introducing a coarse-to-fine architecture that enforces consistency through both implicit and explicit constraints.

SimVG explores the importance of multi-modal understanding for the VG task, proposing a simple yet effective framework. It also adopts a synchronized distillation learning strategy between the teacher and student branches, enhancing the performance of the student branch.

InstanceVG supports instance-level referring segmentation across general scenarios (no/single/multiple targets). It also provides consistent prediction across point, box, and mask.

GC3VG generalizes the C3VG architecture and incorporates UCRM, which implicitly captures region/instance features and explicitly aligns them via an IoU-based relational constraint. The GHA strategy ensures feature-space consistency and boosts the discriminative strength of multi-modal representations.

DRL adopts an end-to-end training and inference paradigm to address common issues in image-retrieval-based UAV self-localization, including complex preprocessing, inherent localization errors, and slow inference.

DenseUAV introduces a real-world sampled dataset for vision-based UAV self-localization and provides a comprehensive benchmark for the task.

FSRA is the first successful application of Transformer models to cross-view geo-localization. It introduces an attention-map-based region partitioning and alignment strategy that alleviates performance degradation caused by viewpoint shifts.