Causal video question answering (QA) has garnered increasing interest, yet existing datasets often lack depth in causal reasoning analysis. To address this gap, we capitalize on the unique properties of cartoons and construct CausalChaos!, a novel, challenging causal Why-QA dataset built upon the iconic "Tom and Jerry" cartoon series. With thoughtful questions and multi-level answers, our dataset contains much longer causal chains embedded in dynamic interactions and visuals, at the same time principles of animation allows animators to create well-defined, unambiguous causal relationships. These factors allow models to solve more challenging, yet well-defined causal relationships. We also introduce hard negative mining, including CausalConfusion version. While models perform well, there is much room for improvement, especially, on open-ended answers. We identify more advanced/explicit causal relationship modeling and joint modeling of vision and language as the immediate areas for future efforts to focus upon. Along with the other complementary datasets, our new challenging dataset will pave the way for these developments in the field. We will release our dataset, codes, and models to help future efforts in this domain.
Zero-shot open-ended inference on untrimmed videos poses a significant challenge, especially when no annotated data is utilized to navigate the inference direction. In this work, we aim to address this underexplored domain by introducing an adaptable framework that efficiently combines both the frozen vision-language (VL) model and off-the-shelf large language model (LLM) for conducting zero-shot open-ended inference tasks without requiring any additional training or fine-tuning. Our comprehensive experiments span various video action datasets for goal inference and action recognition tasks. The results demonstrate the framework's superior performance in goal inference compared to conventional vision-language models in open-ended and close-ended scenarios. Notably, the proposed framework exhibits the capability to generalize effectively to action recognition tasks, underscoring its versatility and potential contributions to advancing the video-based zero-shot understanding.
In this work, we introduce the novel concept of visually Connecting Actions and Their Effects (CATE) in video understanding. CATE can have applications in areas like task planning and learning from demonstration. We propose different CATE-based task formulations, such as action selection and action specification, where video understanding models connect actions and effects at semantic and fine-grained levels. We observe that different formulations produce representations capturing intuitive action properties. We also design various baseline models for action selection and action specification. Despite the intuitive nature of the task, we observe that models struggle, and humans outperform them by a large margin. The study aims to establish a foundation for future efforts, showcasing the flexibility and versatility of connecting actions and effects in video understanding, with the hope of inspiring advanced formulations and models.
Human motion synthesis is a fundamental task in computer animation. Recent methods based on diffusion models or GPT structure demonstrate commendable performance but exhibit drawbacks in terms of slow sampling speeds and error accumulation. In this paper, we propose \emph{Motion Flow Matching}, a novel generative model designed for human motion generation featuring efficient sampling and effectiveness in motion editing applications. Our method reduces the sampling complexity from thousand steps in previous diffusion models to just ten steps, while achieving comparable performance in text-to-motion and action-to-motion generation benchmarks. Noticeably, our approach establishes a new state-of-the-art Fr\'echet Inception Distance on the KIT-ML dataset. What is more, we tailor a straightforward motion editing paradigm named \emph{sampling trajectory rewriting} leveraging the ODE-style generative models and apply it to various editing scenarios including motion prediction, motion in-between prediction, motion interpolation, and upper-body editing. Our code will be released.
In this paper, we study multimodal coreference resolution, specifically where a longer descriptive text, i.e., a narration is paired with an image. This poses significant challenges due to fine-grained image-text alignment, inherent ambiguity present in narrative language, and unavailability of large annotated training sets. To tackle these challenges, we present a data efficient semi-supervised approach that utilizes image-narration pairs to resolve coreferences and narrative grounding in a multimodal context. Our approach incorporates losses for both labeled and unlabeled data within a cross-modal framework. Our evaluation shows that the proposed approach outperforms strong baselines both quantitatively and qualitatively, for the tasks of coreference resolution and narrative grounding.
Situation Recognition is the task of generating a structured summary of what is happening in an image using an activity verb and the semantic roles played by actors and objects. In this task, the same activity verb can describe a diverse set of situations as well as the same actor or object category can play a diverse set of semantic roles depending on the situation depicted in the image. Hence model needs to understand the context of the image and the visual-linguistic meaning of semantic roles. Therefore, we leverage the CLIP foundational model that has learned the context of images via language descriptions. We show that deeper-and-wider multi-layer perceptron (MLP) blocks obtain noteworthy results for the situation recognition task by using CLIP image and text embedding features and it even outperforms the state-of-the-art CoFormer, a Transformer-based model, thanks to the external implicit visual-linguistic knowledge encapsulated by CLIP and the expressive power of modern MLP block designs. Motivated by this, we design a cross-attention-based Transformer using CLIP visual tokens that model the relation between textual roles and visual entities. Our cross-attention-based Transformer known as ClipSitu XTF outperforms existing state-of-the-art by a large margin of 14.1% on semantic role labelling (value) for top-1 accuracy using imSitu dataset. We will make the code publicly available.
While VideoQA Transformer models demonstrate competitive performance on standard benchmarks, the reasons behind their success remain unclear. Do these models jointly capture and leverage the rich multimodal structures and dynamics from video and text? Or are they merely exploiting shortcuts to achieve high scores? We analyze this with $\textit{QUAG}$ (QUadrant AveraGe), a lightweight and non-parametric probe that systematically ablates the model's coupled multimodal understanding during inference. Surprisingly, QUAG reveals that the models manage to maintain high performance even when injected with multimodal sub-optimality. Additionally, even after replacing self-attention in multimodal fusion blocks with "QUAG-attention", a simplistic and less-expressive variant of self-attention, the models maintain high performance. This means that current VideoQA benchmarks and their metrics do not penalize shortcuts that discount joint multimodal understanding. Motivated by this, we propose the $\textit{CLAVI}$ (Counterfactual in LAnguage and VIdeo) benchmark, a diagnostic dataset for benchmarking coupled multimodal understanding in VideoQA through counterfactuals. CLAVI consists of temporal questions and videos that are augmented to curate balanced counterfactuals in language and video domains. Hence, it incentivizes, and identifies the reliability of learnt multimodal representations. We evaluate CLAVI and find that models achieve high performance on multimodal shortcut instances, but have very poor performance on the counterfactuals. Hence, we position CLAVI as a litmus test to identify, diagnose and improve the sub-optimality of learnt multimodal VideoQA representations which the current benchmarks are unable to assess.
Humans have the natural ability to recognize actions even if the objects involved in the action or the background are changed. Humans can abstract away the action from the appearance of the objects and their context which is referred to as compositionality of actions. Compositional action recognition deals with imparting human-like compositional generalization abilities to action-recognition models. In this regard, extracting the interactions between humans and objects forms the basis of compositional understanding. These interactions are not affected by the appearance biases of the objects or the context. But the context provides additional cues about the interactions between things and stuff. Hence we need to infuse context into the human-object interactions for compositional action recognition. To this end, we first design a spatial-temporal interaction encoder that captures the human-object (things) interactions. The encoder learns the spatio-temporal interaction tokens disentangled from the background context. The interaction tokens are then infused with contextual information from the video tokens to model the interactions between things and stuff. The final context-infused spatio-temporal interaction tokens are used for compositional action recognition. We show the effectiveness of our interaction-centric approach on the compositional Something-Else dataset where we obtain a new state-of-the-art result of 83.8% top-1 accuracy outperforming recent important object-centric methods by a significant margin. Our approach of explicit human-object-stuff interaction modeling is effective even for standard action recognition datasets such as Something-Something-V2 and Epic-Kitchens-100 where we obtain comparable or better performance than state-of-the-art.
Visual Abductive Reasoning (VAR) is an emerging vision-language (VL) topic where the model needs to retrieve/generate a likely textual hypothesis from a visual input (image or part of an image) using backward reasoning based on prior knowledge or commonsense. Unlike in conventional VL retrieval or captioning tasks, where entities of texts appear in the image, in abductive inferences, the relevant facts about inferences are not directly visible in the input images. Besides, the inferences are causally relevant to regional visual hints and vary with the latter. Existing works highlight visual parts from a global background with specific prompt tuning techniques (e.g., colorful prompt tuning) on top of foundation models, like CLIP. However, these methods uniformly patchify "regional hints" and "global context" at the same granularity level and may lose fine-grained visual details significant for abductive reasoning. To tackle this, we propose a simple yet effective Regional Prompt Tuning, which encodes "regional visual hints" and "global contexts" separately at fine and coarse-grained levels. Specifically, our model explicitly upsamples, then patchify local hints to get fine-grained regional prompts. These prompts are concatenated with coarse-grained contextual tokens from whole images. We also equip our model with a new Dual-Contrastive Loss to regress the visual feature simultaneously toward features of factual description (a.k.a. clue text) and plausible hypothesis (abductive inference text) during training. Extensive experiments on the Sherlock dataset demonstrate that our fully fine-tuned RGP/RGPs with Dual-Contrastive Loss significantly outperforms previous SOTAs, achieving the 1 rank on abductive reasoning leaderboards among all submissions, under all metrics (e.g., P@1$_{i->t}$: RGPs 38.78 vs CPT-CLIP 33.44, higher=better). We would open-source our codes for further research.
Coreference resolution aims at identifying words and phrases which refer to same entity in a text, a core tool in natural language processing. In this paper, we propose a novel task, resolving coreferences in multimodal data, long-form textual descriptions of visual scenes. Most existing image-text datasets only contain short sentences without coreferent expressions, or coreferences are not annotated. To this end, we first introduce a new dataset, Flickr30k-Coref in which coreference chains and bounding box localization of these chains are annotated. We propose a new technique that learns to identify coreference chains through weakly supervised grounding from image-text pairs and a regularization using prior linguistic knowledge. Our model yields large performance gains over prior work in coreference resolution and weakly supervised grounding of long-form text descriptions.