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.

A wide array of sequence models are built on a framework modeled after Transformers, comprising alternating sequence mixer and channel mixer layers. This paper studies a unifying *matrix mixer* view of sequence mixers that can be conceptualized as a linear map on the input sequence. This framework encompasses a broad range of well-known sequence models, including the self-attention of Transformers as well as recent strong alternatives such as structured state space models (SSMs), and allows understanding downstream characteristics such as efficiency and expressivity through properties of their structured matrix class. We identify a key axis of matrix parameterizations termed *sequence alignment*, which increases the flexibility and performance of matrix mixers, providing insights into the strong performance of Transformers and recent SSMs such as Mamba. Furthermore, the matrix mixer framework offers a systematic approach to developing sequence mixers with desired properties, allowing us to develop several new sub-quadratic sequence models. In particular, we propose a natural bidirectional extension of the Mamba model (**Hydra**), parameterized as a *quasiseparable matrix mixer*, which demonstrates superior performance over other sequence models including Transformers on non-causal tasks. As a drop-in replacement for attention layers, \name outperforms BERT by 0.8 points on the GLUE benchmark and ViT by 2% Top-1 accuracy on ImageNet.

The handling of long videos with complex and occluded sequences has recently emerged as a new challenge in the video instance segmentation (VIS) community. However, existing methods have limitations in addressing this challenge. We argue that the biggest bottleneck in current approaches is the discrepancy between training and inference. To effectively bridge this gap, we propose a Generalized framework for VIS, namely GenVIS, that achieves state-of-the-art performance on challenging benchmarks without designing complicated architectures or requiring extra post-processing. The key contribution of GenVIS is the learning strategy, which includes a query-based training pipeline for sequential learning with a novel target label assignment. Additionally, we introduce a memory that effectively acquires information from previous states. Thanks to the new perspective, which focuses on building relationships between separate frames or clips, GenVIS can be flexibly executed in both online and semi-online manner. We evaluate our approach on popular VIS benchmarks, achieving state-of-the-art results on YouTube-VIS 2019/2021/2022 and Occluded VIS (OVIS). Notably, we greatly outperform the state-of-the-art on the long VIS benchmark (OVIS), improving 5.6 AP with ResNet-50 backbone. Code is available at https://github.com/miranheo/GenVIS.

For online video instance segmentation (VIS), fully utilizing the information from previous frames in an efficient manner is essential for real-time applications. Most previous methods follow a two-stage approach requiring additional computations such as RPN and RoIAlign, and do not fully exploit the available information in the video for all subtasks in VIS. In this paper, we propose a novel single-stage framework for online VIS built based on the grid structured feature representation. The grid-based features allow us to employ fully convolutional networks for real-time processing, and also to easily reuse and share features within different components. We also introduce cooperatively operating modules that aggregate information from available frames, in order to enrich the features for all subtasks in VIS. Our design fully takes advantage of previous information in a grid form for all tasks in VIS in an efficient way, and we achieved the new state-of-the-art accuracy (38.6 AP and 36.9 AP) and speed (40.0 FPS) on YouTube-VIS 2019 and 2021 datasets among online VIS methods. The code is available at https://github.com/SuHoHan95/VISOLQ

Recently, both long-tailed recognition and object tracking have made great advances individually. TAO benchmark presented a mixture of the two, long-tailed object tracking, in order to further reflect the aspect of the real-world. To date, existing solutions have adopted detectors showing robustness in long-tailed distributions, which derive per-frame results. Then, they used tracking algorithms that combine the temporally independent detections to finalize tracklets. However, as the approaches did not take temporal changes in scenes into account, inconsistent classification results in videos led to low overall performance. In this paper, we present a set classifier that improves accuracy of classifying tracklets by aggregating information from multiple viewpoints contained in a tracklet. To cope with sparse annotations in videos, we further propose augmentation of tracklets that can maximize data efficiency. The set classifier is plug-and-playable to existing object trackers, and highly improves the performance of long-tailed object tracking. By simply attaching our method to QDTrack on top of ResNet-101, we achieve the new state-of-the-art, 19.9% and 15.7%$TrackAP_{50}$on TAO validation and test sets, respectively. Our code is available at this link11https://github.com/sukjunhwang/setclassifier.

We introduce a novel paradigm for offline Video Instance Segmentation (VIS), based on the hypothesis that explicit object-oriented information can be a strong clue for understanding the context of the entire sequence. To this end, we propose VITA, a simple structure built on top of an off-the-shelf Transformer-based image instance segmentation model. Specifically, we use an image object detector as a means of distilling object-specific contexts into object tokens. VITA accomplishes video-level understanding by associating frame-level object tokens without using spatio-temporal backbone features. By effectively building relationships between objects using the condensed information, VITA achieves the state-of-the-art on VIS benchmarks with a ResNet-50 backbone: 49.8 AP, 45.7 AP on YouTube-VIS 2019 & 2021, and 19.6 AP on OVIS. Moreover, thanks to its object token-based structure that is disjoint from the backbone features, VITA shows several practical advantages that previous offline VIS methods have not explored - handling long and high-resolution videos with a common GPU, and freezing a frame-level detector trained on image domain. Code is available at the link.

Panoptic segmentation, which is a novel task of unifying instance segmentation and semantic segmentation, has attracted a lot of attention lately. However, most of the previous methods are composed of multiple pathways with each pathway specialized to a designated segmentation task. In this paper, we propose to resolve panoptic segmentation in single-shot by integrating the execution flows. With the integrated pathway, a unified feature map called Panoptic-Feature is generated, which includes the information of both things and stuffs. Panoptic-Feature becomes more sophisticated by auxiliary problems that guide to cluster pixels that belong to the same instance and differentiate between objects of different classes. A collection of convolutional filters, where each filter represents either a thing or stuff, is applied to Panoptic-Feature at once, materializing the single-shot panoptic segmentation. Taking the advantages of both top-down and bottom-up approaches, our method, named SPINet, enjoys high efficiency and accuracy on major panoptic segmentation benchmarks: COCO and Cityscapes.

We propose a novel end-to-end solution for video instance segmentation (VIS) based on transformers. Recently, the per-clip pipeline shows superior performance over per-frame methods leveraging richer information from multiple frames. However, previous per-clip models require heavy computation and memory usage to achieve frame-to-frame communications, limiting practicality.In this work, we propose Inter-frame Communication Transformers (IFC), which significantly reduces the overhead for information-passing between frames by efficiently encoding the context within the input clip.Specifically, we propose to utilize concise memory tokens as a means of conveying information as well as summarizing each frame scene.The features of each frame are enriched and correlated with other frames through exchange of information between the precisely encoded memory tokens.We validate our method on the latest benchmark sets and achieved state-of-the-art performance (AP 42.6 on YouTube-VIS 2019 val set using the offline inference) while having a considerably fast runtime (89.4 FPS). Our method can also be applied to near-online inference for processing a video in real-time with only a small delay.The code is available at https://github.com/sukjunhwang/IFC