MRBench: A Comprehensive Benchmark for Human Motion-Text Retrieval

Pretraining Body Part Representations for Text-Motion Retrieval

Text-motion retrieval has gained increasing research attention, yet several critical challenges remain such as data scarcity, limited fine-grained matching capabilities, and inadequate evaluation protocols. To address these issues, we propose POP-TMR which pretrains body part representations for fine-grained text-motion retrieval. Our approach leverages large-scale human motion datasets to pretrain a spatio-temporal transformer-based motion encoder, enabling more generalizable motion features. In addition to matching global motion and text representations, we propose a local branch to capture detailed body part features for enhancing spatial-aware cross-modal alignment. To improve evaluation, we introduce HumanML3D+, an enhanced benchmark that provides accurate positive annotations for text queries and includes text descriptions at varying levels of detail, enabling more systematic performance assessment. Extensive experiments on KIT-ML, HumanML3D and HumanML3D+ benchmarks demonstrate that POP-TMR outperforms state-of-the-art methods. Furthermore, we showcase its effectiveness in additional downstream applications, including text-to-motion generation evaluation, human interaction recognition and zero-shot moment retrieval. Data, code and pretrained model are publicly available at https://lin-kayla.github.io/POP_TMR/.

FineMoLA: Towards Fine-Grained Motion-Language Alignment from Clip-Level Supervision

Text-conditioned human motion generation has made rapid progress with the emergence of large-scale motion--language datasets. However, even datasets with rich long-form descriptions typically provide supervision only at the clip level, without explicit temporal correspondence between motion frames and language. This limits fine-grained motion--text grounding and temporally precise generation. We propose FineMoLA, a weakly supervised framework that learns fine-grained frame--phrase correspondence directly from clip-level annotations. Our method first segments long-form descriptions into action-bearing phrases, and then formulates motion--language alignment as an optimal transport problem, which naturally models many-to-many relations between motion frames and text under global constraints. With entropic regularization and Sinkhorn iterations, FineMoLA efficiently infers pseudo frame-level alignments without human labeling. Experiments on SnapMoGen demonstrate that the learned alignments outperform baselines in motion--text grounding.

MoGenVD: A Motion-Centered Quality Assessment Benchmark for Text-to-Video Generation

Multimodal Video Understanding: A Capability-Based Survey of Alignment, Expression, and Reasoning

Multimodal video understanding (MVU) has emerged as a fast-growing research frontier, driven by major advances in video-language pre-training and large multimodal models over the past decade. MVU aims to synergistically integrate visual, audio and textual modalities to interpret complex video semantics, supporting widespread downstream tasks including cross-modal retrieval, dense captioning, video question answering, event analysis and intelligent assistance. Despite the rapid proliferation of specialized MVU models, the community still lacks a unified capability-centric framework to systematically clarify the hierarchical competency architecture and evolutionary trajectory of state-of-the-art approaches. To address this issue, this paper presents a structured, comprehensive survey of the latest MVU progress, establishing a novel three-tier taxonomy that categorizes existing studies into cross-modal alignment, multi-granularity semantic expression and multimodal reasoning. Along this pipeline, we further systematically synthesize core modality fusion strategies, mainstream benchmark datasets and standardized evaluation protocols. Through a fine-grained analysis of representative published results, we highlight the critical impact of inconsistent evaluation settings, cross-experiment comparability bottlenecks and inherent methodological trade-offs between performance and efficiency. Finally, we identify and dissect three key open challenges: ultra-long video scalability, performance degradation from modality noise and missing data, and factual reliability risks in generative MVU systems. This capability-oriented systematic reference clarifies the methodological evolution logic of MVU, and provides actionable guidance for developing next-generation robust, high-performance multimodal video understanding systems.

PHA-Net: Prototype-based hierarchical alignment network for text-video retrieval

With the emergence of large-scale image-text pre-training models, e.g., CLIP, text-video retrieval has experienced substantial advances in recent years. Existing best-performing methods involve aligning cross-modal semantics at individual, local, and global levels simultaneously, raising concerns about the intrinsic semantic mismatch between concise texts and rich videos. A canonical approach is to integrate multiple language-video attention modules into the hierarchical framework while this paradigm only optimizes visual representations with prohibitive computational costs. In this paper, we propose a new prototype-based hierarchical alignment network (PHA-Net) to align individual/local/global level representations across modalities. Concretely, we introduce multiple modality-shared prototypes as the bridge to efficiently optimize text and video representations for cross-modal alignment. Then, we argue that the imbalanced semantic distribution in clustered tokens may undermine retrieval performance, as tokens with weak semantics are of little interest. To reduce the impact of these tokens, a proposed prototype-supported token merge module is responsible for enhancing tokens with strong semantics and suppressing others with weak semantics via prototype semantics guidance. Moreover, we devise a prototype contrastive loss to encourage textual and visual prototypes to focus on different semantic information. The idea of this auxiliary loss is to ensure higher similarity between textual and visual prototypes from the same prototype than those from different prototypes. Extensive experiments on four benchmarks confirm the effectiveness of our PHA-Net, which achieves significant improvements in the sum of all recalls on MSR-VTT (8.8%), ActivityNet (19.2%), VATEX (0.7%), and Charades (4.9%). Code is available at https://github.com/JingXiaolun/PHA-Net.

Training-Free Video Corpus Moment Retrieval via Synergistic Collaboration and Adaptive Calibration

Video Corpus Moment Retrieval (VCMR) is pivotal to multimodal understanding. However, existing methods rely heavily on large-scale annotated data, which limits their generalization and scalability. To address this issue, we propose a training-free VCMR framework, termed Synergistic Collaboration and Adaptive Calibration (SCAC), enabling effective semantic parsing and precise temporal localization without parameter updates. SCAC introduces a Query Event Chain Generation module that leverages large language models to transform complex textual queries into structured event chains, while a Video Event Chain Generation module represents videos as semantically coherent event chains through subtitle segmentation and keyframe aggregation. Built on these structured representations, SCAC performs Event-Chain-Based Cross-Modal Retrieval with mean-variance joint scoring to suppress local mismatches and reinforce global consistency. During localization, a Synergy-Calibration Mechanism dynamically refines temporal boundaries via profit-setback feedback. Extensive experiments show that SCAC achieves comparable or superior results to supervised counterparts under training-free conditions, demonstrating strong cross-modal generalization and adaptive capability. The code of our method is available at https://github.com/cyanlll/SCAC