As short-form funny videos on social networks are gaining popularity, it becomes demanding for AI models to understand them for better communication with humans. Unfortunately, previous video humor datasets target specific domains, such as speeches or sitcoms, and mostly focus on verbal cues. We curate a user-generated dataset of 10K multimodal funny videos from YouTube, called ExFunTube. Using a video filtering pipeline with GPT-3.5, we verify both verbal and visual elements contributing to humor. After filtering, we annotate each video with timestamps and text explanations for funny moments. Our ExFunTube is unique over existing datasets in that our videos cover a wide range of domains with various types of humor that necessitate a multimodal understanding of the content. Also, we develop a zero-shot video-to-text prompting to maximize video humor understanding of large language models (LLMs). With three different evaluation methods using automatic scores, rationale quality experiments, and human evaluations, we show that our prompting significantly improves LLMs' ability for humor explanation.
相關內容
The strong temporal consistency of surveillance video enables compelling compression performance with traditional methods, but downstream vision applications operate on decoded image frames with a high data rate. Since it is not straightforward for applications to extract information on temporal redundancy from the compressed video representations, we propose a novel system which conveys temporal redundancy within a sparse decompressed representation. We leverage a video representation framework called ADDER to transcode framed videos to sparse, asynchronous intensity samples. We introduce mechanisms for content adaptation, lossy compression, and asynchronous forms of classical vision algorithms. We evaluate our system on the VIRAT surveillance video dataset, and we show a median 43.7% speed improvement in FAST feature detection compared to OpenCV. We run the same algorithm as OpenCV, but only process pixels that receive new asynchronous events, rather than process every pixel in an image frame. Our work paves the way for upcoming neuromorphic sensors and is amenable to future applications with spiking neural networks.
Incremental computation aims to compute more efficiently on changed input by reusing previously computed results. We give a high-level overview of works on incremental computation, and highlight the essence underlying all of them, which we call incrementalization -- the discrete counterpart of differentiation in calculus. We review the gist of a systematic method for incrementalization, and a systematic method centered around it, called Iterate-Incrementalize-Implement, for program design and optimization, as well as algorithm design and optimization. At a meta-level, with historical contexts and for future directions, we stress the power of high-level data, control, and module abstractions in developing new and better algorithms and programs as well as their precise complexities.
Current training of motion style transfer systems relies on consistency losses across style domains to preserve contents, hindering its scalable application to a large number of domains and private data. Recent image transfer works show the potential of independent training on each domain by leveraging implicit bridging between diffusion models, with the content preservation, however, limited to simple data patterns. We address this by imposing biased sampling in backward diffusion while maintaining the domain independence in the training stage. We construct the bias from the source domain keyframes and apply them as the gradient of content constraints, yielding a framework with keyframe manifold constraint gradients (KMCGs). Our validation demonstrates the success of training separate models to transfer between as many as ten dance motion styles. Comprehensive experiments find a significant improvement in preserving motion contents in comparison to baseline and ablative diffusion-based style transfer models. In addition, we perform a human study for a subjective assessment of the quality of generated dance motions. The results validate the competitiveness of KMCGs.
The Segment Anything Model (SAM) is the first foundation model for general image segmentation. It has achieved impressive results on various natural image segmentation tasks. However, medical image segmentation (MIS) is more challenging because of the complex modalities, fine anatomical structures, uncertain and complex object boundaries, and wide-range object scales. To fully validate SAM's performance on medical data, we collected and sorted 53 open-source datasets and built a large medical segmentation dataset with 18 modalities, 84 objects, 125 object-modality paired targets, 1050K 2D images, and 6033K masks. We comprehensively analyzed different models and strategies on the so-called COSMOS 1050K dataset. Our findings mainly include the following: 1) SAM showed remarkable performance in some specific objects but was unstable, imperfect, or even totally failed in other situations. 2) SAM with the large ViT-H showed better overall performance than that with the small ViT-B. 3) SAM performed better with manual hints, especially box, than the Everything mode. 4) SAM could help human annotation with high labeling quality and less time. 5) SAM was sensitive to the randomness in the center point and tight box prompts, and may suffer from a serious performance drop. 6) SAM performed better than interactive methods with one or a few points, but will be outpaced as the number of points increases. 7) SAM's performance correlated to different factors, including boundary complexity, intensity differences, etc. 8) Finetuning the SAM on specific medical tasks could improve its average DICE performance by 4.39% and 6.68% for ViT-B and ViT-H, respectively. We hope that this comprehensive report can help researchers explore the potential of SAM applications in MIS, and guide how to appropriately use and develop SAM.
We present W.A.L.T, a transformer-based approach for photorealistic video generation via diffusion modeling. Our approach has two key design decisions. First, we use a causal encoder to jointly compress images and videos within a unified latent space, enabling training and generation across modalities. Second, for memory and training efficiency, we use a window attention architecture tailored for joint spatial and spatiotemporal generative modeling. Taken together these design decisions enable us to achieve state-of-the-art performance on established video (UCF-101 and Kinetics-600) and image (ImageNet) generation benchmarks without using classifier free guidance. Finally, we also train a cascade of three models for the task of text-to-video generation consisting of a base latent video diffusion model, and two video super-resolution diffusion models to generate videos of $512 \times 896$ resolution at $8$ frames per second.
Existing approaches to video understanding, mainly designed for short videos from a third-person perspective, are limited in their applicability in certain fields, such as robotics. In this paper, we delve into open-ended question-answering (QA) in long, egocentric videos, which allows individuals or robots to inquire about their own past visual experiences. This task presents unique challenges, including the complexity of temporally grounding queries within extensive video content, the high resource demands for precise data annotation, and the inherent difficulty of evaluating open-ended answers due to their ambiguous nature. Our proposed approach tackles these challenges by (i) integrating query grounding and answering within a unified model to reduce error propagation; (ii) employing large language models for efficient and scalable data synthesis; and (iii) introducing a close-ended QA task for evaluation, to manage answer ambiguity. Extensive experiments demonstrate the effectiveness of our method, which also achieves state-of-the-art performance on the QAEgo4D and Ego4D-NLQ benchmarks. We plan to publicly release the codes, model, and constructed datasets for future research.
Diffusion models achieve remarkable quality in image generation, but at a cost. Iterative denoising requires many time steps to produce high fidelity images. We argue that the denoising process is crucially limited by an accumulation of the reconstruction error due to an initial inaccurate reconstruction of the target data. This leads to lower quality outputs, and slower convergence. To address this issue, we propose compensation sampling to guide the generation towards the target domain. We introduce a compensation term, implemented as a U-Net, which adds negligible computation overhead during training and, optionally, inference. Our approach is flexible and we demonstrate its application in unconditional generation, face inpainting, and face de-occlusion using benchmark datasets CIFAR-10, CelebA, CelebA-HQ, FFHQ-256, and FSG. Our approach consistently yields state-of-the-art results in terms of image quality, while accelerating the denoising process to converge during training by up to an order of magnitude.
Temporal action localization aims to identify the boundaries and categories of actions in videos, such as scoring a goal in a football match. Single-frame supervision has emerged as a labor-efficient way to train action localizers as it requires only one annotated frame per action. However, it often suffers from poor performance due to the lack of precise boundary annotations. To address this issue, we propose a visual analysis method that aligns similar actions and then propagates a few user-provided annotations (e.g. , boundaries, category labels) to similar actions via the generated alignments. Our method models the alignment between actions as a heaviest path problem and the annotation propagation as a quadratic optimization problem. As the automatically generated alignments may not accurately match the associated actions and could produce inaccurate localization results, we develop a storyline visualization to explain the localization results of actions and their alignments. This visualization facilitates users in correcting wrong localization results and misalignments. The corrections are then used to improve the localization results of other actions. The effectiveness of our method in improving localization performance is demonstrated through quantitative evaluation and a case study.
Generative models are now capable of producing highly realistic images that look nearly indistinguishable from the data on which they are trained. This raises the question: if we have good enough generative models, do we still need datasets? We investigate this question in the setting of learning general-purpose visual representations from a black-box generative model rather than directly from data. Given an off-the-shelf image generator without any access to its training data, we train representations from the samples output by this generator. We compare several representation learning methods that can be applied to this setting, using the latent space of the generator to generate multiple "views" of the same semantic content. We show that for contrastive methods, this multiview data can naturally be used to identify positive pairs (nearby in latent space) and negative pairs (far apart in latent space). We find that the resulting representations rival those learned directly from real data, but that good performance requires care in the sampling strategy applied and the training method. Generative models can be viewed as a compressed and organized copy of a dataset, and we envision a future where more and more "model zoos" proliferate while datasets become increasingly unwieldy, missing, or private. This paper suggests several techniques for dealing with visual representation learning in such a future. Code is released on our project page: //ali-design.github.io/GenRep/
Language model pre-training has proven to be useful in learning universal language representations. As a state-of-the-art language model pre-training model, BERT (Bidirectional Encoder Representations from Transformers) has achieved amazing results in many language understanding tasks. In this paper, we conduct exhaustive experiments to investigate different fine-tuning methods of BERT on text classification task and provide a general solution for BERT fine-tuning. Finally, the proposed solution obtains new state-of-the-art results on eight widely-studied text classification datasets.
小貼士
登錄享
注冊地址： 北京市海淀區羊坊店路18號2幢3層301-191