Kumara Kahatapitiya

Michael S. Ryoo, Keerthana Gopalakrishnan, Kumara Kahatapitiya et al.

We propose Token Turing Machines (TTM), a sequential, autoregressive Transformer model with memory for real-world sequential visual understanding. Our model is inspired by the seminal Neural Turing Machine, and has an external memory consisting of a set of tokens which summarise the previous history (i.e., frames). This memory is efficiently addressed, read and written using a Transformer as the processing unit/controller at each step. The model's memory module ensures that a new observation will only be processed with the contents of the memory (and not the entire history), meaning that it can efficiently process long sequences with a bounded computational cost at each step. We show that TTM outperforms other alternatives, such as other Transformer models designed for long sequences and recurrent neural networks, on two real-world sequential visual understanding tasks: online temporal activity detection from videos and vision-based robot action policy learning. Code is publicly available at: https://github.com/google-research/scenic/tree/main/scenic/projects/token_turing

Kumara Kahatapitiya, Anurag Arnab, Arsha Nagrani et al.

Vision-Language models (VLMs) have excelled in the image-domain -- especially in zero-shot settings -- thanks to the availability of vast pretraining data (i.e., paired image-text samples). However for videos, such paired data is not as abundant. Therefore, video-VLMs are usually designed by adapting pretrained image-VLMs to the video-domain, instead of training from scratch. All such recipes rely on augmenting visual embeddings with temporal information (i.e., image $\rightarrow$ video), often keeping text embeddings unchanged or even being discarded. In this paper, we argue the contrary, that better video-VLMs can be designed by focusing more on augmenting text, rather than visual information. More specifically, we introduce Video-conditioned Text Representations (VicTR): a form of text embeddings optimized w.r.t. visual embeddings, creating a more-flexible contrastive latent space. Our model can further make use of freely-available semantic information, in the form of visually-grounded auxiliary text (e.g. object or scene information). We evaluate our model on few-shot, zero-shot (HMDB-51, UCF-101), short-form (Kinetics-400) and long-form (Charades) activity recognition benchmarks, showing strong performance among video-VLMs.

Jongwoo Park, Kumara Kahatapitiya, Donghyun Kim et al.

Vision Transformers (ViTs) have recently become the state-of-the-art across many computer vision tasks. In contrast to convolutional networks (CNNs), ViTs enable global information sharing even within shallow layers of a network, i.e., among high-resolution features. However, this perk was later overlooked with the success of pyramid architectures such as Swin Transformer, which show better performance-complexity trade-offs. In this paper, we present a simple and efficient add-on component (termed GrafT) that considers global dependencies and multi-scale information throughout the network, in both high- and low-resolution features alike. It has the flexibility of branching out at arbitrary depths and shares most of the parameters and computations of the backbone. GrafT shows consistent gains over various well-known models which includes both hybrid and pure Transformer types, both homogeneous and pyramid structures, and various self-attention methods. In particular, it largely benefits mobile-size models by providing high-level semantics. On the ImageNet-1k dataset, GrafT delivers +3.9%, +1.4%, and +1.9% top-1 accuracy improvement to DeiT-T, Swin-T, and MobileViT-XXS, respectively. Our code and models will be made available.

Kumara Kahatapitiya, Kanchana Ranasinghe, Jongwoo Park et al.

Language has become a prominent modality in computer vision with the rise of LLMs. Despite supporting long context-lengths, their effectiveness in handling long-term information gradually declines with input length. This becomes critical, especially in applications such as long-form video understanding. In this paper, we introduce a Language Repository (LangRepo) for LLMs, that maintains concise and structured information as an interpretable (i.e., all-textual) representation. Our repository is updated iteratively based on multi-scale video chunks. We introduce write and read operations that focus on pruning redundancies in text, and extracting information at various temporal scales. The proposed framework is evaluated on zero-shot visual question-answering benchmarks including EgoSchema, NExT-QA, IntentQA and NExT-GQA, showing state-of-the-art performance at its scale. Our code is available at https://github.com/kkahatapitiya/LangRepo.

Kanchana Ranasinghe, Xiang Li, Kumara Kahatapitiya et al.

Large Language Models (LLMs) have allowed recent LLM-based approaches to achieve excellent performance on long-video understanding benchmarks. We investigate how extensive world knowledge and strong reasoning skills of underlying LLMs influence this strong performance. Surprisingly, we discover that LLM-based approaches can yield surprisingly good accuracy on long-video tasks with limited video information, sometimes even with no video specific information. Building on this, we explore injecting video-specific information into an LLM-based framework. We utilize off-the-shelf vision tools to extract three object-centric information modalities from videos, and then leverage natural language as a medium for fusing this information. Our resulting Multimodal Video Understanding (MVU) framework demonstrates state-of-the-art performance across multiple video understanding benchmarks. Strong performance also on robotics domain tasks establish its strong generality. Code: https://github.com/kahnchana/mvu

Zhaochong An, Menglin Jia, Haonan Qiu et al.

Storytelling in real-world videos often unfolds through multiple shots -- discontinuous yet semantically connected clips that together convey a coherent narrative. However, existing multi-shot video generation (MSV) methods struggle to effectively model long-range cross-shot context, as they rely on limited temporal windows or single keyframe conditioning, leading to degraded performance under complex narratives. In this work, we propose OneStory, enabling global yet compact cross-shot context modeling for consistent and scalable narrative generation. OneStory reformulates MSV as a next-shot generation task, enabling autoregressive shot synthesis while leveraging pretrained image-to-video (I2V) models for strong visual conditioning. We introduce two key modules: a Frame Selection module that constructs a semantically-relevant global memory based on informative frames from prior shots, and an Adaptive Conditioner that performs importance-guided patchification to generate compact context for direct conditioning. We further curate a high-quality multi-shot dataset with referential captions to mirror real-world storytelling patterns, and design effective training strategies under the next-shot paradigm. Finetuned from a pretrained I2V model on our curated 60K dataset, OneStory achieves state-of-the-art narrative coherence across diverse and complex scenes in both text- and image-conditioned settings, enabling controllable and immersive long-form video storytelling.

Xiang Li, Cristina Mata, Jongwoo Park et al.

Vision Language Models (VLMs) have recently been leveraged to generate robotic actions, forming Vision-Language-Action (VLA) models. However, directly adapting a pretrained VLM for robotic control remains challenging, particularly when constrained by a limited number of robot demonstrations. In this work, we introduce LLaRA: Large Language and Robotics Assistant, a framework that formulates robot action policy as visuo-textual conversations and enables an efficient transfer of a pretrained VLM into a powerful VLA, motivated by the success of visual instruction tuning in Computer Vision. First, we present an automated pipeline to generate conversation-style instruction tuning data for robots from existing behavior cloning datasets, aligning robotic actions with image pixel coordinates. Further, we enhance this dataset in a self-supervised manner by defining six auxiliary tasks, without requiring any additional action annotations. We show that a VLM finetuned with a limited amount of such datasets can produce meaningful action decisions for robotic control. Through experiments across multiple simulated and real-world tasks, we demonstrate that LLaRA achieves state-of-the-art performance while preserving the generalization capabilities of large language models. The code, datasets, and pretrained models are available at https://github.com/LostXine/LLaRA.

Jongwoo Park, Kanchana Ranasinghe, Kumara Kahatapitiya et al.

Long-form videos that span across wide temporal intervals are highly information redundant and contain multiple distinct events or entities that are often loosely related. Therefore, when performing long-form video question answering (LVQA), all information necessary to generate a correct response can often be contained within a small subset of frames. Recent literature leverage large language models (LLMs) in LVQA benchmarks, achieving exceptional performance, while relying on vision language models (VLMs) to convert all visual content within videos into natural language. Such VLMs often independently caption a large number of frames uniformly sampled from long videos, which is not efficient and can mostly be redundant. Motivated by this inefficiency, we propose LVNet, a modular and training-free framework featuring a novel Hierarchical Keyframe Selector (HKS) that efficiently selects a minimal set of informative frames tailored to each question. LVNet's modularity allows easy integration with existing approaches for more efficient LVQA. We achieve state-of-the-art performance among similarly configured models across four benchmark LVQA datasets: EgoSchema, NExT-QA, IntentQA, VideoMME. The code can be found at https://github.com/jongwoopark7978/LVNet

Kumara Kahatapitiya, Michael S. Ryoo

Modeling visual data as tokens (i.e., image patches) using attention mechanisms, feed-forward networks or convolutions has been highly effective in recent years. Such methods usually have a common pipeline: a tokenization method, followed by a set of layers/blocks for information mixing, both within and among tokens. When image patches are converted into tokens, they are often flattened, discarding the spatial structure within each patch. As a result, any processing that follows (eg: multi-head self-attention) may fail to recover and/or benefit from such information. In this paper, we argue that models can have significant gains when spatial structure is preserved during tokenization, and is explicitly used during the mixing stage. We propose two key contributions: (1) Structure-aware Tokenization and, (2) Structure-aware Mixing, both of which can be combined with existing models with minimal effort. We introduce a family of models (SWAT), showing improvements over the likes of DeiT, MLP-Mixer and Swin Transformer, across multiple benchmarks including ImageNet classification and ADE20K segmentation. Our code is available at https://github.com/kkahatapitiya/SWAT.

Kumara Kahatapitiya, Zhou Ren, Haoxiang Li et al.

Temporal Activity Detection aims to predict activity classes per frame, in contrast to video-level predictions in Activity Classification (i.e., Activity Recognition). Due to the expensive frame-level annotations required for detection, the scale of detection datasets is limited. Thus, commonly, previous work on temporal activity detection resorts to fine-tuning a classification model pretrained on large-scale classification datasets (e.g., Kinetics-400). However, such pretrained models are not ideal for downstream detection, due to the disparity between the pretraining and the downstream fine-tuning tasks. In this work, we propose a novel 'weakly-guided self-supervised' pretraining method for detection. We leverage weak labels (classification) to introduce a self-supervised pretext task (detection) by generating frame-level pseudo labels, multi-action frames, and action segments. Simply put, we design a detection task similar to downstream, on large-scale classification data, without extra annotations. We show that the models pretrained with the proposed weakly-guided self-supervised detection task outperform prior work on multiple challenging activity detection benchmarks, including Charades and MultiTHUMOS. Our extensive ablations further provide insights on when and how to use the proposed models for activity detection. Code is available at https://github.com/kkahatapitiya/SSDet.

Jinghuan Shang, Kumara Kahatapitiya, Xiang Li et al.

Reinforcement Learning (RL) can be considered as a sequence modeling task: given a sequence of past state-action-reward experiences, an agent predicts a sequence of next actions. In this work, we propose State-Action-Reward Transformer (StARformer) for visual RL, which explicitly models short-term state-action-reward representations (StAR-representations), essentially introducing a Markovian-like inductive bias to improve long-term modeling. Our approach first extracts StAR-representations by self-attending image state patches, action, and reward tokens within a short temporal window. These are then combined with pure image state representations -- extracted as convolutional features, to perform self-attention over the whole sequence. Our experiments show that StARformer outperforms the state-of-the-art Transformer-based method on image-based Atari and DeepMind Control Suite benchmarks, in both offline-RL and imitation learning settings. StARformer is also more compliant with longer sequences of inputs. Our code is available at https://github.com/elicassion/StARformer.

Kumara Kahatapitiya, Michael S. Ryoo

In this paper, we introduce Coarse-Fine Networks, a two-stream architecture which benefits from different abstractions of temporal resolution to learn better video representations for long-term motion. Traditional Video models process inputs at one (or few) fixed temporal resolution without any dynamic frame selection. However, we argue that, processing multiple temporal resolutions of the input and doing so dynamically by learning to estimate the importance of each frame can largely improve video representations, specially in the domain of temporal activity localization. To this end, we propose (1) Grid Pool, a learned temporal downsampling layer to extract coarse features, and, (2) Multi-stage Fusion, a spatio-temporal attention mechanism to fuse a fine-grained context with the coarse features. We show that our method outperforms the state-of-the-arts for action detection in public datasets including Charades with a significantly reduced compute and memory footprint. The code is available at https://github.com/kkahatapitiya/Coarse-Fine-Networks

Kumara Kahatapitiya, Haozhe Liu, Sen He et al.

Generating temporally-consistent high-fidelity videos can be computationally expensive, especially over longer temporal spans. More-recent Diffusion Transformers (DiTs) -- despite making significant headway in this context -- have only heightened such challenges as they rely on larger models and heavier attention mechanisms, resulting in slower inference speeds. In this paper, we introduce a training-free method to accelerate video DiTs, termed Adaptive Caching (AdaCache), which is motivated by the fact that "not all videos are created equal": meaning, some videos require fewer denoising steps to attain a reasonable quality than others. Building on this, we not only cache computations through the diffusion process, but also devise a caching schedule tailored to each video generation, maximizing the quality-latency trade-off. We further introduce a Motion Regularization (MoReg) scheme to utilize video information within AdaCache, essentially controlling the compute allocation based on motion content. Altogether, our plug-and-play contributions grant significant inference speedups (e.g. up to 4.7x on Open-Sora 720p - 2s video generation) without sacrificing the generation quality, across multiple video DiT baselines.

Haozhe Liu, Shikun Liu, Zijian Zhou et al.

We introduce MarDini, a new family of video diffusion models that integrate the advantages of masked auto-regression (MAR) into a unified diffusion model (DM) framework. Here, MAR handles temporal planning, while DM focuses on spatial generation in an asymmetric network design: i) a MAR-based planning model containing most of the parameters generates planning signals for each masked frame using low-resolution input; ii) a lightweight generation model uses these signals to produce high-resolution frames via diffusion de-noising. MarDini's MAR enables video generation conditioned on any number of masked frames at any frame positions: a single model can handle video interpolation (e.g., masking middle frames), image-to-video generation (e.g., masking from the second frame onward), and video expansion (e.g., masking half the frames). The efficient design allocates most of the computational resources to the low-resolution planning model, making computationally expensive but important spatio-temporal attention feasible at scale. MarDini sets a new state-of-the-art for video interpolation; meanwhile, within few inference steps, it efficiently generates videos on par with those of much more expensive advanced image-to-video models.

Kumara Kahatapitiya, Adil Karjauv, Davide Abati et al.

Diffusion-based video editing have reached impressive quality and can transform either the global style, local structure, and attributes of given video inputs, following textual edit prompts. However, such solutions typically incur heavy memory and computational costs to generate temporally-coherent frames, either in the form of diffusion inversion and/or cross-frame attention. In this paper, we conduct an analysis of such inefficiencies, and suggest simple yet effective modifications that allow significant speed-ups whilst maintaining quality. Moreover, we introduce Object-Centric Diffusion, to fix generation artifacts and further reduce latency by allocating more computations towards foreground edited regions, arguably more important for perceptual quality. We achieve this by two novel proposals: i) Object-Centric Sampling, decoupling the diffusion steps spent on salient or background regions and spending most on the former, and ii) Object-Centric Token Merging, which reduces cost of cross-frame attention by fusing redundant tokens in unimportant background regions. Both techniques are readily applicable to a given video editing model without retraining, and can drastically reduce its memory and computational cost. We evaluate our proposals on inversion-based and control-signal-based editing pipelines, and show a latency reduction up to 10x for a comparable synthesis quality. Project page: qualcomm-ai-research.github.io/object-centric-diffusion.

Rui Dai, Srijan Das, Kumara Kahatapitiya et al.

Action detection is an essential and challenging task, especially for densely labelled datasets of untrimmed videos. The temporal relation is complex in those datasets, including challenges like composite action, and co-occurring action. For detecting actions in those complex videos, efficiently capturing both short-term and long-term temporal information in the video is critical. To this end, we propose a novel ConvTransformer network for action detection. This network comprises three main components: (1) Temporal Encoder module extensively explores global and local temporal relations at multiple temporal resolutions. (2) Temporal Scale Mixer module effectively fuses the multi-scale features to have a unified feature representation. (3) Classification module is used to learn the instance center-relative position and predict the frame-level classification scores. The extensive experiments on multiple datasets, including Charades, TSU and MultiTHUMOS, confirm the effectiveness of our proposed method. Our network outperforms the state-of-the-art methods on all three datasets.

Dumindu Tissera, Kasun Vithanage, Rukshan Wijesinghe et al.

Learning a particular task from a dataset, samples in which originate from diverse contexts, is challenging, and usually addressed by deepening or widening standard neural networks. As opposed to conventional network widening, multi-path architectures restrict the quadratic increment of complexity to a linear scale. However, existing multi-column/path networks or model ensembling methods do not consider any feature-dependent allocation of parallel resources, and therefore, tend to learn redundant features. Given a layer in a multi-path network, if we restrict each path to learn a context-specific set of features and introduce a mechanism to intelligently allocate incoming feature maps to such paths, each path can specialize in a certain context, reducing the redundancy and improving the quality of extracted features. This eventually leads to better-optimized usage of parallel resources. To do this, we propose inserting feature-dependent cross-connections between parallel sets of feature maps in successive layers. The weighting coefficients of these cross-connections are computed from the input features of the particular layer. Our multi-path networks show improved image recognition accuracy at a similar complexity compared to conventional and state-of-the-art methods for deepening, widening and adaptive feature extracting, in both small and large scale datasets.

Dumindu Tissera, Kumara Kahatapitiya, Rukshan Wijesinghe et al.

Making a single network effectively address diverse contexts---learning the variations within a dataset or multiple datasets---is an intriguing step towards achieving generalized intelligence. Existing approaches of deepening, widening, and assembling networks are not cost effective in general. In view of this, networks which can allocate resources according to the context of the input and regulate flow of information across the network are effective. In this paper, we present Context-Aware Multipath Network (CAMNet), a multi-path neural network with data-dependant routing between parallel tensors. We show that our model performs as a generalized model capturing variations in individual datasets and multiple different datasets, both simultaneously and sequentially. CAMNet surpasses the performance of classification and pixel-labeling tasks in comparison with the equivalent single-path, multi-path, and deeper single-path networks, considering datasets individually, sequentially, and in combination. The data-dependent routing between tensors in CAMNet enables the model to control the flow of information end-to-end, deciding which resources to be common or domain-specific.

Kumara Kahatapitiya, Ranga Rodrigo

Convolutional Neural Networks (CNNs) have achieved state-of-the-art performance in many computer vision tasks over the years. However, this comes at the cost of heavy computation and memory intensive network designs, suggesting potential improvements in efficiency. Convolutional layers of CNNs partly account for such an inefficiency, as they are known to learn redundant features. In this work, we exploit this redundancy, observing it as the correlation between convolutional filters of a layer, and propose an alternative approach to reproduce it efficiently. The proposed 'LinearConv' layer learns a set of orthogonal filters, and a set of coefficients that linearly combines them to introduce a controlled redundancy. We introduce a correlation-based regularization loss to achieve such flexibility over redundancy, and control the number of parameters in turn. This is designed as a plug-and-play layer to conveniently replace a conventional convolutional layer, without any additional changes required in the network architecture or the hyperparameter settings. Our experiments verify that LinearConv models achieve a performance on-par with their counterparts, with almost a 50% reduction in parameters on average, and the same computational requirement and speed at inference.

Kumara Kahatapitiya, Dumindu Tissera, Ranga Rodrigo

Occlusion removal is an interesting application of image enhancement, for which, existing work suggests manually-annotated or domain-specific occlusion removal. No work tries to address automatic occlusion detection and removal as a context-aware generic problem. In this paper, we present a novel methodology to identify objects that do not relate to the image context as occlusions and remove them, reconstructing the space occupied coherently. The proposed system detects occlusions by considering the relation between foreground and background object classes represented as vector embeddings, and removes them through inpainting. We test our system on COCO-Stuff dataset and conduct a user study to establish a baseline in context-aware automatic occlusion removal.