Video large language models (LLMs) achieve strong video understanding by leveraging a large number of spatio-temporal tokens, but suffer from quadratic computational scaling with token count. To address this, we propose a training-free spatio-temporal token merging method, named STTM. Our key insight is to exploit local spatial and temporal redundancy in video data which has been overlooked in prior work. STTM first transforms each frame into multi-granular spatial tokens using a coarse-to-fine search over a quadtree structure, then performs directed pairwise merging across the temporal dimension. This decomposed merging approach outperforms existing token reduction methods across six video QA benchmarks. Notably, STTM achieves a 2$\times$ speed-up with only a 0.5% accuracy drop under a 50% token budget, and a 3$\times$ speed-up with just a 2% drop under a 30% budget. Moreover, STTM is query-agnostic, allowing KV cache reuse across different questions for the same video. The project page is available at https://www.jshyun.me/projects/sttm.

In a retrieval system, simultaneously achieving search accuracy and efficiency is inherently challenging. This challenge is particularly pronounced in partially relevant video retrieval (PRVR), where incorporating more diverse context representations at varying temporal scales for each video enhances accuracy but increases computational and memory costs. To address this dichotomy, we propose a prototypical PRVR framework that encodes diverse contexts within a video into a fixed number of prototypes. We then introduce several strategies to enhance text association and video understanding within the prototypes, along with an orthogonal objective to ensure that the prototypes capture a diverse range of content. To keep the prototypes searchable via text queries while accurately encoding video contexts, we implement cross- and uni-modal reconstruction tasks. The cross-modal reconstruction task aligns the prototypes with textual features within a shared space, while the uni-modal reconstruction task preserves all video contexts during encoding. Additionally, we employ a video mixing technique to provide weak guidance to further align prototypes and associated textual representations. Extensive evaluations on TVR, ActivityNet-Captions, and QVHighlights validate the effectiveness of our approach without sacrificing efficiency.

Most of the existing Video Anomaly Detection (VAD) studies have been conducted within single-domain learning, where training and evaluation are performed on a single dataset. However, the criteria for abnormal events differ across VAD datasets, making it problematic to apply a single-domain model to other domains. In this paper, we propose a new task called Multi-Domain learning forVAD (MDVAD) to explore various real-world abnormal events using multiple datasets for a general model. MDVAD involves training on datasets from multiple domains simultaneously, and we experimentally observe that Abnormal Conflicts between domains hinder learning and generalization. The task aims to address two key objectives: (i) better distinguishing between general normal and abnormal events across multiple domains, and (ii) being aware of ambiguous abnormal conflicts. This paper is the first to tackle abnormal conflict issue and introduces a new benchmark, baselines, and evaluation protocols for MDVAD. As baselines, we propose a framework with Null(Angular)-Multiple Instance Learning and an Abnormal Conflict classifier. Through experiments on a MDVAD benchmark composed of six VAD datasets and using four different evaluation protocols, we reveal abnormal conflicts and demonstrate that the proposed baseline effectively handles these conflicts, showing robustness and adaptability across multiple domains.

Recent self-supervised video representation learning methods focus on maximizing the similarity between multiple augmented views from the same video and largely rely on the quality of generated views. However, most existing methods lack a mechanism to prevent representation learning from bias towards static information in the video. In this paper, we propose frequency augmentation (FreqAug), a spatio-temporal data augmentation method in the frequency domain for video representation learning. FreqAug stochastically removes specific frequency components from the video so that learned representation captures essential features more from the remaining information for various downstream tasks. Specifically, FreqAug pushes the model to focus more on dynamic features rather than static features in the video via dropping spatial or temporal low-frequency components. To verify the generality of the proposed method, we experiment with FreqAug on multiple self-supervised learning frameworks along with standard augmentations. Transferring the improved representation to five video action recognition and two temporal action localization downstream tasks shows consistent improvements over baselines.

Temporal action detection aims to predict the time intervals and the classes of action instances in the video. Despite the promising performance, existing two-stream models exhibit slow inference speed due to their reliance on computationally expensive optical flow. In this paper, we introduce a decomposed cross-modal distillation framework to build a strong RGB-based detector by transferring knowledge of the motion modality. Specifically, instead of direct distillation, we propose to separately learn RGB and motion representations, which are in turn combined to perform action localization. The dual-branch design and the asymmetric training objectives enable effective motion knowledge transfer while preserving RGB information intact. In addition, we introduce a local attentive fusion to better exploit the multimodal complementarity. It is designed to preserve the local discriminability of the features that is important for action localization. Extensive experiments on the benchmarks verify the effectiveness of the proposed method in enhancing RGB-based action detectors. Notably, our framework is agnostic to backbones and detection heads, bringing consistent gains across different model combinations.

Most existing multi-person tracking approaches rely on appearance based re-identification (re-ID) to resolve fragmented tracklets. However, simply using appearance information could be insufficient for videos containing severe pose changes, such as sports or dance videos. With the goal of learning pose-invariant representations, we propose an end-to-end deep learning framework Sparse-Temporal ReID Network. Our proposed network not only realizes human pose disentanglement in an image recovery manner, but also makes efficient linkages between the identical subjects via a unique Sparse temporal identity sampling technique across time steps. Experimental results demonstrate the effectiveness of our proposed method on both multi-view re-ID benchmarks and our newly collected dance video dataset DanceReID1.