temporal segmentation

This work addresses the limited parallelizability of the classical dynamic time warping (DTW) algorithm, which suffers from quadratic time and memory complexity. The authors propose Segmental DTW, a novel approach that decomposes global sequence alignment into local subsequence DTW computations that can be executed in parallel, followed by a segment-level dynamic programming step to integrate the partial alignments. This method achieves near-full parallelism while preserving alignment accuracy comparable to standard DTW. Theoretical analysis and empirical evaluation on Chopin Mazurka audio alignment tasks demonstrate that one variant of the proposed method outperforms existing approaches in both computational efficiency and alignment performance.

This work addresses the inefficiency of traditional Dynamic Time Warping (DTW) in aligning long sequences due to its inherently serial computation, which hinders effective GPU parallelization. The study presents the first systematic exploration of GPU-oriented parallel alternatives to DTW, introducing four novel algorithms. The first three employ rectangular block-based approximations to accelerate computation, while the fourth, termed ParDTW, achieves exact alignment through diagonal-wise parallelization. ParDTW integrates block matrix processing with a diagonal scheduling strategy, preserving full alignment accuracy while delivering 15–100× speedup over existing methods on long sequences. This breakthrough substantially overcomes the performance limitations of conventional DTW, establishing ParDTW as an efficient and practical solution for large-scale sequence alignment tasks.

Existing video generation models struggle to model overlapping events because their temporal representations are limited to discrete time steps, lacking the ability to capture time intervals and concurrent relationships. This work proposes Temporal Interval Encoding (TIE), which for the first time treats time intervals as first-class primitives within a RoPE-compatible bilinear attention mechanism. By leveraging interval integration and duration invariance principles, TIE derives a closed-form sinc-based solution that enables interval-aware generation without altering the standard attention interface. Evaluated on the OmniEvents dataset, TIE improves the human-verified temporal constraint satisfaction rate from 77.34% to 96.03% and reduces temporal boundary error from 0.261 seconds to 0.073 seconds, all while preserving the original visual quality of the DiT architecture.

To address ambiguous event boundaries and frame-level misalignment in Partially Relevant Video Retrieval (PRVR), this paper proposes the first unequal-length event modeling framework. Methodologically: (1) a Progressive Grouped Video Segmentation (PGVS) module is designed to enable text-guided, dynamic boundary partitioning; (2) Cross-modal Attention-guided Event Refinement (CAER) is introduced to jointly incorporate temporal dependencies and inter-frame semantic similarity, thereby enhancing fine-grained text-video alignment. Evaluated on two standard PRVR benchmarks, our approach achieves state-of-the-art performance, significantly improving localization accuracy and retrieval recall for partially relevant segments. Experimental results validate the effectiveness of modeling events as semantically coherent, boundary-precise, and variable-length units—marking a departure from conventional fixed-length or rigidly segmented paradigms.

Human motion videos in complex backgrounds often violate the union-of-subspaces (UoS) assumption, hindering effective non-overlapping human motion segmentation (HMS). Method: This paper proposes a temporally consistent structured clustering framework for HMS. Its core innovation is the first introduction of a temporal rate reduction criterion to drive subspace alignment, enabling dynamic representations to naturally conform to the UoS structure and overcoming limitations of conventional assumptions. The framework further integrates information-theoretic rate reduction optimization, end-to-end structured representation learning, adaptive affinity graph construction, and spectral clustering. Contribution/Results: The method is compatible with diverse feature extractors and achieves state-of-the-art performance across five standard HMS benchmarks. It significantly improves action boundary localization accuracy—particularly in cluttered background scenarios—demonstrating robustness and generalizability.

This work addresses the limited ability of existing motion-language models to achieve fine-grained understanding and precise control over body parts, which hinders advancements in animation and interactive applications. To overcome this, we propose MotionMERGE, the first unified multi-granularity motion framework that supports fine-grained language-guided motion generation. By explicitly modeling motion patterns across body parts and temporal dimensions, MotionMERGE enables high-fidelity motion understanding, editing, and synthesis. Key innovations include a joint pretraining strategy—Reasoning-Aware Granularity-Synergy—that aligns representations across granularities while incorporating motion reasoning, a chain-of-thought inference mechanism, and MotionFineEdit, the first large-scale dataset annotated with spatiotemporal correction instructions and reasoning chains. Experiments demonstrate that our approach significantly outperforms current models across multiple tasks and exhibits exceptional zero-shot generalization, advancing the frontier of fine-grained human-motion interaction.

This work addresses the challenges faced by existing video temporal localization methods—such as high computational cost, sparse attention, and loss of spatial detail—in modeling temporal dynamics. The authors propose a novel time-to-space transformation paradigm that restructures a sequence of video frames into a structured 2D grid image using overlapping sliding windows, thereby preserving fine-grained spatial information while explicitly encoding global temporal relationships. By integrating composite textual timestamps with a vision-language foundation model architecture, the approach unifies local attention mechanisms with global temporal awareness. The method achieves state-of-the-art performance across multiple standard video temporal localization benchmarks, significantly improving both localization accuracy and temporal understanding.

This work addresses the challenge of imprecise temporal localization in zero-shot video moment retrieval, which arises from a mismatch in semantic granularity between textual queries and visual content. To resolve this issue, the authors propose a training-free, granularity-aware alignment framework that achieves precise cross-modal semantic alignment across multiple levels of granularity. The approach leverages multi-granularity query rewriting and query-aware video caption generation to bridge the semantic gap without requiring any model training. As the first zero-shot method to explicitly incorporate granularity awareness, this framework establishes new state-of-the-art results on three major benchmarks—QVHighlights, Charades-STA, and ActivityNet-Captions—with a notable 3.23% absolute improvement in mAP@avg on QVHighlights.

This work addresses the limitations of existing video temporal localization methods, which rely on endpoint prediction and struggle to explicitly model the validity, duration, and interval-level semantics of temporal segments. The authors propose TimePLE, a novel approach that reformulates the task as an interval-native joint distribution prediction, directly modeling valid intervals in a position–duration squared space and generating continuous boundaries via a single <|TIMESPAN|> latent state. Key innovations include a geometrically consistent position–duration representation, a duration-aware coordinate correction mechanism, and implicit interval-anchor alignment through a vision-language model. The study also introduces a large-scale, human-verified dataset for precise full-event alignment. Evaluated across four benchmarks, TimePLE achieves an average mIoU of 58.9%, substantially outperforming endpoint-based baselines, with notable gains on short and medium-duration events.

This work addresses the quadratic computational bottleneck of self-attention in video referring segmentation caused by dense spatiotemporal tokens. It introduces state space models to this task for the first time, proposing a linear-complexity approach for temporal context modeling. By leveraging decoupled bidirectional spatial scanning and causal temporal scanning mechanisms, together with a segmentation-target-driven adaptive thresholding strategy, the method achieves hierarchical and dynamic spatiotemporal token compression. This design preserves semantic completeness while enabling zero-shot streaming inference. Experiments demonstrate that the proposed approach surpasses non-compressed baselines across multiple benchmarks, achieving an 85% reduction in tokens and a 1.8× speedup in inference, effectively balancing efficiency and accuracy.