Yulei Niu

Beier Zhu, Yulei Niu, Yucheng Han et al.

Thanks to the large pre-trained vision-language models (VLMs) like CLIP, we can craft a zero-shot classifier by "prompt", e.g., the confidence score of an image being "[CLASS]" can be obtained by using the VLM provided similarity measure between the image and the prompt sentence "a photo of a [CLASS]". Therefore, prompt shows a great potential for fast adaptation of VLMs to downstream tasks if we fine-tune the prompt-based similarity measure. However, we find a common failure that improper fine-tuning may not only undermine the prompt's inherent prediction for the task-related classes, but also for other classes in the VLM vocabulary. Existing methods still address this problem by using traditional anti-overfitting techniques such as early stopping and data augmentation, which lack a principled solution specific to prompt. We present Prompt-aligned Gradient, dubbed ProGrad, to prevent prompt tuning from forgetting the the general knowledge learned from VLMs. In particular, ProGrad only updates the prompt whose gradient is aligned (or non-conflicting) to the "general direction", which is represented as the gradient of the KL loss of the pre-defined prompt prediction. Extensive experiments demonstrate the stronger few-shot generalization ability of ProGrad over state-of-the-art prompt tuning methods. Codes are available at https://github.com/BeierZhu/Prompt-align.

Xinting Hu, Yulei Niu, Chunyan Miao et al.

Semi-Supervised Learning (SSL) is fundamentally a missing label problem, in which the label Missing Not At Random (MNAR) problem is more realistic and challenging, compared to the widely-adopted yet naive Missing Completely At Random assumption where both labeled and unlabeled data share the same class distribution. Different from existing SSL solutions that overlook the role of "class" in causing the non-randomness, e.g., users are more likely to label popular classes, we explicitly incorporate "class" into SSL. Our method is three-fold: 1) We propose Class-Aware Propensity (CAP) that exploits the unlabeled data to train an improved classifier using the biased labeled data. 2) To encourage rare class training, whose model is low-recall but high-precision that discards too many pseudo-labeled data, we propose Class-Aware Imputation (CAI) that dynamically decreases (or increases) the pseudo-label assignment threshold for rare (or frequent) classes. 3) Overall, we integrate CAP and CAI into a Class-Aware Doubly Robust (CADR) estimator for training an unbiased SSL model. Under various MNAR settings and ablations, our method not only significantly outperforms existing baselines but also surpasses other label bias removal SSL methods. Please check our code at: https://github.com/JoyHuYY1412/CADR-FixMatch.

Yulei Niu, Long Chen, Chang Zhou et al.

Knowledge distillation (KD) is essentially a process of transferring a teacher model's behavior, e.g., network response, to a student model. The network response serves as additional supervision to formulate the machine domain, which uses the data collected from the human domain as a transfer set. Traditional KD methods hold an underlying assumption that the data collected in both human domain and machine domain are both independent and identically distributed (IID). We point out that this naive assumption is unrealistic and there is indeed a transfer gap between the two domains. Although the gap offers the student model external knowledge from the machine domain, the imbalanced teacher knowledge would make us incorrectly estimate how much to transfer from teacher to student per sample on the non-IID transfer set. To tackle this challenge, we propose Inverse Probability Weighting Distillation (IPWD) that estimates the propensity score of a training sample belonging to the machine domain, and assigns its inverse amount to compensate for under-represented samples. Experiments on CIFAR-100 and ImageNet demonstrate the effectiveness of IPWD for both two-stage distillation and one-stage self-distillation.

Beier Zhu, Yulei Niu, Saeil Lee et al.

We present a new paradigm for fine-tuning large-scale visionlanguage pre-trained models on downstream task, dubbed Prompt Regularization (ProReg). Different from traditional fine-tuning which easily overfits to the downstream task data, ProReg uses the prediction by prompting the pretrained model to regularize the fine-tuning. The motivation is: by prompting the large model "a photo of a [CLASS]", the fil-lin answer is only dependent on the pretraining encyclopedic knowledge while independent of the task data distribution, which is usually biased. Specifically, given a training sample prediction during fine-tuning, we first calculate its KullbackLeibler loss of the prompt prediction and Cross-Entropy loss of the ground-truth label, and then combine them with a proposed sample-wise adaptive trade-off weight, which automatically adjusts the transfer between the pretrained and downstream domains. On various out-of-distribution benchmarks, we show the consistently strong performance of ProReg compared with conventional fine-tuning, zero-shot prompt, prompt tuning, and other state-of-the-art methods.

Yaojie Shen, Xinyao Wang, Yulei Niu et al.

Preference Optimization (PO), is gaining popularity as an alternative choice of Proximal Policy Optimization (PPO) for aligning Large Language Models (LLMs). Recent research on aligning LLMs iteratively with synthetic or partially synthetic data shows promising results in scaling up PO training for both academic settings and proprietary trained models such as Llama3. Despite its success, our study shows that the length exploitation issue present in PO is even more severe in Iterative Preference Optimization (IPO) due to the iterative nature of the process. In this work, we study iterative preference optimization with synthetic data. We share the findings and analysis along the way of building the iterative preference optimization pipeline. More specifically, we discuss the length exploitation issue during iterative preference optimization and propose our training objective for iterative preference optimization, namely Agreement-aware Iterative Preference Optimization (AIPO). To demonstrate the effectiveness of our method, we conduct comprehensive experiments and achieve state-of-the-art performance on MT-Bench, AlpacaEval 2.0, and Arena-Hard. Our implementation and model checkpoints will be made available at https://github.com/bytedance/AIPO.

Hung-Ting Su, Yulei Niu, Xudong Lin et al.

Causal Video Question Answering (CVidQA) queries not only association or temporal relations but also causal relations in a video. Existing question synthesis methods pre-trained question generation (QG) systems on reading comprehension datasets with text descriptions as inputs. However, QG models only learn to ask association questions (e.g., ``what is someone doing...'') and result in inferior performance due to the poor transfer of association knowledge to CVidQA, which focuses on causal questions like ``why is someone doing ...''. Observing this, we proposed to exploit causal knowledge to generate question-answer pairs, and proposed a novel framework, Causal Knowledge Extraction from Language Models (CaKE-LM), leveraging causal commonsense knowledge from language models to tackle CVidQA. To extract knowledge from LMs, CaKE-LM generates causal questions containing two events with one triggering another (e.g., ``score a goal'' triggers ``soccer player kicking ball'') by prompting LM with the action (soccer player kicking ball) to retrieve the intention (to score a goal). CaKE-LM significantly outperforms conventional methods by 4% to 6% of zero-shot CVidQA accuracy on NExT-QA and Causal-VidQA datasets. We also conduct comprehensive analyses and provide key findings for future research.

Jiawei Ma, Yulei Niu, Jincheng Xu et al.

Generalized few-shot object detection aims to achieve precise detection on both base classes with abundant annotations and novel classes with limited training data. Existing approaches enhance few-shot generalization with the sacrifice of base-class performance, or maintain high precision in base-class detection with limited improvement in novel-class adaptation. In this paper, we point out the reason is insufficient Discriminative feature learning for all of the classes. As such, we propose a new training framework, DiGeo, to learn Geometry-aware features of inter-class separation and intra-class compactness. To guide the separation of feature clusters, we derive an offline simplex equiangular tight frame (ETF) classifier whose weights serve as class centers and are maximally and equally separated. To tighten the cluster for each class, we include adaptive class-specific margins into the classification loss and encourage the features close to the class centers. Experimental studies on two few-shot benchmark datasets (VOC, COCO) and one long-tail dataset (LVIS) demonstrate that, with a single model, our method can effectively improve generalization on novel classes without hurting the detection of base classes.

Long Chen, Yulei Niu, Brian Chen et al.

Given a long untrimmed video and natural language queries, video grounding (VG) aims to temporally localize the semantically-aligned video segments. Almost all existing VG work holds two simple but unrealistic assumptions: 1) All query sentences can be grounded in the corresponding video. 2) All query sentences for the same video are always at the same semantic scale. Unfortunately, both assumptions make today's VG models fail to work in practice. For example, in real-world multimodal assets (eg, news articles), most of the sentences in the article can not be grounded in their affiliated videos, and they typically have rich hierarchical relations (ie, at different semantic scales). To this end, we propose a new challenging grounding task: Weakly-Supervised temporal Article Grounding (WSAG). Specifically, given an article and a relevant video, WSAG aims to localize all ``groundable'' sentences to the video, and these sentences are possibly at different semantic scales. Accordingly, we collect the first WSAG dataset to facilitate this task: YouwikiHow, which borrows the inherent multi-scale descriptions in wikiHow articles and plentiful YouTube videos. In addition, we propose a simple but effective method DualMIL for WSAG, which consists of a two-level MIL loss and a single-/cross- sentence constraint loss. These training objectives are carefully designed for these relaxed assumptions. Extensive ablations have verified the effectiveness of DualMIL.

Zhen Wang, Long Chen, Wenbo Ma et al.

Given an image and a reference caption, the image caption editing task aims to correct the misalignment errors and generate a refined caption. However, all existing caption editing works are implicit models, ie, they directly produce the refined captions without explicit connections to the reference captions. In this paper, we introduce a new task: Explicit Caption Editing (ECE). ECE models explicitly generate a sequence of edit operations, and this edit operation sequence can translate the reference caption into a refined one. Compared to the implicit editing, ECE has multiple advantages: 1) Explainable: it can trace the whole editing path. 2) Editing Efficient: it only needs to modify a few words. 3) Human-like: it resembles the way that humans perform caption editing, and tries to keep original sentence structures. To solve this new task, we propose the first ECE model: TIger. TIger is a non-autoregressive transformer-based model, consisting of three modules: Tagger_del, Tagger_add, and Inserter. Specifically, Tagger_del decides whether each word should be preserved or not, Tagger_add decides where to add new words, and Inserter predicts the specific word for adding. To further facilitate ECE research, we propose two new ECE benchmarks by re-organizing two existing datasets, dubbed COCO-EE and Flickr30K-EE, respectively. Extensive ablations on both two benchmarks have demonstrated the effectiveness of TIger.

Hammad A. Ayyubi, Christopher Thomas, Lovish Chum et al.

Events describe happenings in our world that are of importance. Naturally, understanding events mentioned in multimedia content and how they are related forms an important way of comprehending our world. Existing literature can infer if events across textual and visual (video) domains are identical (via grounding) and thus, on the same semantic level. However, grounding fails to capture the intricate cross-event relations that exist due to the same events being referred to on many semantic levels. For example, in Figure 1, the abstract event of "war" manifests at a lower semantic level through subevents "tanks firing" (in video) and airplane "shot" (in text), leading to a hierarchical, multimodal relationship between the events. In this paper, we propose the task of extracting event hierarchies from multimodal (video and text) data to capture how the same event manifests itself in different modalities at different semantic levels. This reveals the structure of events and is critical to understanding them. To support research on this task, we introduce the Multimodal Hierarchical Events (MultiHiEve) dataset. Unlike prior video-language datasets, MultiHiEve is composed of news video-article pairs, which makes it rich in event hierarchies. We densely annotate a part of the dataset to construct the test benchmark. We show the limitations of state-of-the-art unimodal and multimodal baselines on this task. Further, we address these limitations via a new weakly supervised model, leveraging only unannotated video-article pairs from MultiHiEve. We perform a thorough evaluation of our proposed method which demonstrates improved performance on this task and highlight opportunities for future research.

Andrew Lu, Xudong Lin, Yulei Niu et al.

Understanding event relationships in videos requires a model to understand the underlying structures of events (i.e. the event type, the associated argument roles, and corresponding entities) and factual knowledge for reasoning. Structural symbolic representation (SSR) based methods directly take event types and associated argument roles/entities as inputs to perform reasoning. However, the state-of-the-art video event-relation prediction system shows the necessity of using continuous feature vectors from input videos; existing methods based solely on SSR inputs fail completely, even when given oracle event types and argument roles. In this paper, we conduct an extensive empirical analysis to answer the following questions: 1) why SSR-based method failed; 2) how to understand the evaluation setting of video event relation prediction properly; 3) how to uncover the potential of SSR-based methods. We first identify suboptimal training settings as causing the failure of previous SSR-based video event prediction models. Then through qualitative and quantitative analysis, we show how evaluation that takes only video as inputs is currently unfeasible, as well as the reliance on oracle event information to obtain an accurate evaluation. Based on these findings, we propose to further contextualize the SSR-based model to an Event-Sequence Model and equip it with more factual knowledge through a simple yet effective way of reformulating external visual commonsense knowledge bases into an event-relation prediction pretraining dataset. The resultant new state-of-the-art model eventually establishes a 25% Macro-accuracy performance boost.

Hung-Ting Su, Ya-Ching Hsu, Xudong Lin et al.

Large language models (LLMs) equipped with chain-of-thoughts (CoT) prompting have shown significant multi-step reasoning capabilities in factual content like mathematics, commonsense, and logic. However, their performance in narrative reasoning, which demands greater abstraction capabilities, remains unexplored. This study utilizes tropes in movie synopses to assess the abstract reasoning abilities of state-of-the-art LLMs and uncovers their low performance. We introduce a trope-wise querying approach to address these challenges and boost the F1 score by 11.8 points. Moreover, while prior studies suggest that CoT enhances multi-step reasoning, this study shows CoT can cause hallucinations in narrative content, reducing GPT-4's performance. We also introduce an Adversarial Injection method to embed trope-related text tokens into movie synopses without explicit tropes, revealing CoT's heightened sensitivity to such injections. Our comprehensive analysis provides insights for future research directions.

Ying Zhou, Xinyao Wang, Yulei Niu et al.

Recent advancements in large language models (LLMs) have significantly enhanced their knowledge and generative capabilities, leading to a surge of interest in leveraging LLMs for high-quality data synthesis. However, synthetic data generation via prompting LLMs remains challenging due to LLMs' limited understanding of target data distributions and the complexity of prompt engineering, especially for structured formatted data. To address these issues, we introduce DiffLM, a controllable data synthesis framework based on variational autoencoder (VAE), which further (1) leverages diffusion models to reserve more information of original distribution and format structure in the learned latent distribution and (2) decouples the learning of target distribution knowledge from the LLM's generative objectives via a plug-and-play latent feature injection module. As we observed significant discrepancies between the VAE's latent representations and the real data distribution, the latent diffusion module is introduced into our framework to learn a fully expressive latent distribution. Evaluations on seven real-world datasets with structured formatted data (i.e., Tabular, Code, and Tool data) demonstrate that DiffLM generates high-quality data, with performance on downstream tasks surpassing that of real data by 2%-7% in certain cases. Data and code are available at https://github.com/bytedance/DiffLM.

Xiaoying Xing, Chia-Wen Kuo, Li Fuxin et al.

Large Vision-Language Models (LVLMs) have shown promising performance in vision-language understanding and reasoning tasks. However, their visual understanding behaviors remain underexplored. A fundamental question arises: to what extent do LVLMs rely on visual input, and which image regions contribute to their responses? It is non-trivial to interpret the free-form generation of LVLMs due to their complicated visual architecture (e.g., multiple encoders and multi-resolution) and variable-length outputs. In this paper, we extend existing heatmap visualization methods (e.g., iGOS++) to support LVLMs for open-ended visual question answering. We propose a method to select visually relevant tokens that reflect the relevance between generated answers and input image. Furthermore, we conduct a comprehensive analysis of state-of-the-art LVLMs on benchmarks designed to require visual information to answer. Our findings offer several insights into LVLM behavior, including the relationship between focus region and answer correctness, differences in visual attention across architectures, and the impact of LLM scale on visual understanding. The code and data are available at https://github.com/bytedance/LVLM_Interpretation.

Bingxuan Li, Yiming Cui, Yicheng He et al.

Sound effects build an essential layer of multimodal storytelling, shaping the emotional atmosphere and the narrative semantics of videos. Despite recent advancement in video-text-to-audio (VT2A), the current formulation faces three key limitations: First, an imbalance between visual and textual conditioning that leads to visual dominance; Second, the absence of a concrete definition for fine-grained controllable generation; Third, weak instruction understanding and following, as existing datasets rely on brief categorical tags. To address these limitations, we introduce EchoFoley, a new task designed for video-grounded sound generation with both event level local control and hierarchical semantic control. Our symbolic representation for sounding events specifies when, what, and how each sound is produced within a video or instruction, enabling fine-grained controls like sound generation, insertion, and editing. To support this task, we construct EchoFoley-6k, a large-scale, expert-curated benchmark containing over 6,000 video-instruction-annotation triplets. Building upon this foundation, we propose EchoVidia a sounding-event-centric agentic generation framework with slow-fast thinking strategy. Experiments show that EchoVidia surpasses recent VT2A models by 40.7% in controllability and 12.5% in perceptual quality.

Fangyi Chen, Yaojie Shen, Lu Xu et al.

Precise, object-aware control over visual content is essential for advanced image editing and compositional generation. Yet, most existing approaches operate on entire images holistically, limiting the ability to isolate and manipulate individual scene elements. In contrast, layered representations, where scenes are explicitly separated into objects, environmental context, and visual effects, provide a more intuitive and structured framework for interpreting and editing visual content. To bridge this gap and enable both compositional understanding and controllable editing, we introduce the Referring Layer Decomposition (RLD) task, which predicts complete RGBA layers from a single RGB image, conditioned on flexible user prompts, such as spatial inputs (e.g., points, boxes, masks), natural language descriptions, or combinations thereof. At the core is the RefLade, a large-scale dataset comprising 1.11M image-layer-prompt triplets produced by our scalable data engine, along with 100K manually curated, high-fidelity layers. Coupled with a perceptually grounded, human-preference-aligned automatic evaluation protocol, RefLade establishes RLD as a well-defined and benchmarkable research task. Building on this foundation, we present RefLayer, a simple baseline designed for prompt-conditioned layer decomposition, achieving high visual fidelity and semantic alignment. Extensive experiments show our approach enables effective training, reliable evaluation, and high-quality image decomposition, while exhibiting strong zero-shot generalization capabilities.

Beier Zhu, Yulei Niu, Xian-Sheng Hua et al.

We address the overlooked unbiasedness in existing long-tailed classification methods: we find that their overall improvement is mostly attributed to the biased preference of tail over head, as the test distribution is assumed to be balanced; however, when the test is as imbalanced as the long-tailed training data -- let the test respect Zipf's law of nature -- the tail bias is no longer beneficial overall because it hurts the head majorities. In this paper, we propose Cross-Domain Empirical Risk Minimization (xERM) for training an unbiased model to achieve strong performances on both test distributions, which empirically demonstrates that xERM fundamentally improves the classification by learning better feature representation rather than the head vs. tail game. Based on causality, we further theoretically explain why xERM achieves unbiasedness: the bias caused by the domain selection is removed by adjusting the empirical risks on the imbalanced domain and the balanced but unseen domain. Codes are available at https://github.com/BeierZhu/xERM.

Kaifeng Gao, Long Chen, Yulei Niu et al.

Today's VidSGG models are all proposal-based methods, i.e., they first generate numerous paired subject-object snippets as proposals, and then conduct predicate classification for each proposal. In this paper, we argue that this prevalent proposal-based framework has three inherent drawbacks: 1) The ground-truth predicate labels for proposals are partially correct. 2) They break the high-order relations among different predicate instances of a same subject-object pair. 3) VidSGG performance is upper-bounded by the quality of the proposals. To this end, we propose a new classification-then-grounding framework for VidSGG, which can avoid all the three overlooked drawbacks. Meanwhile, under this framework, we reformulate the video scene graphs as temporal bipartite graphs, where the entities and predicates are two types of nodes with time slots, and the edges denote different semantic roles between these nodes. This formulation takes full advantage of our new framework. Accordingly, we further propose a novel BIpartite Graph based SGG model: BIG. It consists of a classification stage and a grounding stage, where the former aims to classify the categories of all the nodes and the edges, and the latter tries to localize the temporal location of each relation instance. Extensive ablations on two VidSGG datasets have attested to the effectiveness of our framework and BIG. Code is available at https://github.com/Dawn-LX/VidSGG-BIG.

Yulei Niu, Kaihua Tang, Hanwang Zhang et al.

VQA models may tend to rely on language bias as a shortcut and thus fail to sufficiently learn the multi-modal knowledge from both vision and language. Recent debiasing methods proposed to exclude the language prior during inference. However, they fail to disentangle the "good" language context and "bad" language bias from the whole. In this paper, we investigate how to mitigate language bias in VQA. Motivated by causal effects, we proposed a novel counterfactual inference framework, which enables us to capture the language bias as the direct causal effect of questions on answers and reduce the language bias by subtracting the direct language effect from the total causal effect. Experiments demonstrate that our proposed counterfactual inference framework 1) is general to various VQA backbones and fusion strategies, 2) achieves competitive performance on the language-bias sensitive VQA-CP dataset while performs robustly on the balanced VQA v2 dataset without any augmented data. The code is available at https://github.com/yuleiniu/cfvqa.

An Zhao, Mingyu Ding, Zhiwu Lu et al.

Existing few-shot learning (FSL) methods make the implicit assumption that the few target class samples are from the same domain as the source class samples. However, in practice this assumption is often invalid -- the target classes could come from a different domain. This poses an additional challenge of domain adaptation (DA) with few training samples. In this paper, the problem of domain-adaptive few-shot learning (DA-FSL) is tackled, which requires solving FSL and DA in a unified framework. To this end, we propose a novel domain-adversarial prototypical network (DAPN) model. It is designed to address a specific challenge in DA-FSL: the DA objective means that the source and target data distributions need to be aligned, typically through a shared domain-adaptive feature embedding space; but the FSL objective dictates that the target domain per class distribution must be different from that of any source domain class, meaning aligning the distributions across domains may harm the FSL performance. How to achieve global domain distribution alignment whilst maintaining source/target per-class discriminativeness thus becomes the key. Our solution is to explicitly enhance the source/target per-class separation before domain-adaptive feature embedding learning in the DAPN, in order to alleviate the negative effect of domain alignment on FSL. Extensive experiments show that our DAPN outperforms the state-of-the-art FSL and DA models, as well as their naïve combinations. The code is available at https://github.com/dingmyu/DAPN.

Jiaxin Qi, Yulei Niu, Jianqiang Huang et al.

This paper unravels the design tricks adopted by us, the champion team MReaL-BDAI, for Visual Dialog Challenge 2019: two causal principles for improving Visual Dialog (VisDial). By "improving", we mean that they can promote almost every existing VisDial model to the state-of-the-art performance on the leader-board. Such a major improvement is only due to our careful inspection on the causality behind the model and data, finding that the community has overlooked two causalities in VisDial. Intuitively, Principle 1 suggests: we should remove the direct input of the dialog history to the answer model, otherwise a harmful shortcut bias will be introduced; Principle 2 says: there is an unobserved confounder for history, question, and answer, leading to spurious correlations from training data. In particular, to remove the confounder suggested in Principle 2, we propose several causal intervention algorithms, which make the training fundamentally different from the traditional likelihood estimation. Note that the two principles are model-agnostic, so they are applicable in any VisDial model. The code is available at https://github.com/simpleshinobu/visdial-principles.

Yulei Niu, Hanwang Zhang, Manli Zhang et al.

Visual dialog is a challenging vision-language task, which requires the agent to answer multi-round questions about an image. It typically needs to address two major problems: (1) How to answer visually-grounded questions, which is the core challenge in visual question answering (VQA); (2) How to infer the co-reference between questions and the dialog history. An example of visual co-reference is: pronouns (\eg, ``they'') in the question (\eg, ``Are they on or off?'') are linked with nouns (\eg, ``lamps'') appearing in the dialog history (\eg, ``How many lamps are there?'') and the object grounded in the image. In this work, to resolve the visual co-reference for visual dialog, we propose a novel attention mechanism called Recursive Visual Attention (RvA). Specifically, our dialog agent browses the dialog history until the agent has sufficient confidence in the visual co-reference resolution, and refines the visual attention recursively. The quantitative and qualitative experimental results on the large-scale VisDial v0.9 and v1.0 datasets demonstrate that the proposed RvA not only outperforms the state-of-the-art methods, but also achieves reasonable recursion and interpretable attention maps without additional annotations. The code is available at \url{https://github.com/yuleiniu/rva}.

Yulei Niu, Wenliang Guo, Long Chen et al.

We study the problem of procedure planning in instructional videos, which aims to make a goal-oriented sequence of action steps given partial visual state observations. The motivation of this problem is to learn a structured and plannable state and action space. Recent works succeeded in sequence modeling of steps with only sequence-level annotations accessible during training, which overlooked the roles of states in the procedures. In this work, we point out that State CHangEs MAtter (SCHEMA) for procedure planning in instructional videos. We aim to establish a more structured state space by investigating the causal relations between steps and states in procedures. Specifically, we explicitly represent each step as state changes and track the state changes in procedures. For step representation, we leveraged the commonsense knowledge in large language models (LLMs) to describe the state changes of steps via our designed chain-of-thought prompting. For state change tracking, we align visual state observations with language state descriptions via cross-modal contrastive learning, and explicitly model the intermediate states of the procedure using LLM-generated state descriptions. Experiments on CrossTask, COIN, and NIV benchmark datasets demonstrate that our proposed SCHEMA model achieves state-of-the-art performance and obtains explainable visualizations.

Ali Zare, Yulei Niu, Hammad Ayyubi et al.

Procedure Planning in instructional videos entails generating a sequence of action steps based on visual observations of the initial and target states. Despite the rapid progress in this task, there remain several critical challenges to be solved: (1) Adaptive procedures: Prior works hold an unrealistic assumption that the number of action steps is known and fixed, leading to non-generalizable models in real-world scenarios where the sequence length varies. (2) Temporal relation: Understanding the step temporal relation knowledge is essential in producing reasonable and executable plans. (3) Annotation cost: Annotating instructional videos with step-level labels (i.e., timestamp) or sequence-level labels (i.e., action category) is demanding and labor-intensive, limiting its generalizability to large-scale datasets. In this work, we propose a new and practical setting, called adaptive procedure planning in instructional videos, where the procedure length is not fixed or pre-determined. To address these challenges, we introduce Retrieval-Augmented Planner (RAP) model. Specifically, for adaptive procedures, RAP adaptively determines the conclusion of actions using an auto-regressive model architecture. For temporal relation, RAP establishes an external memory module to explicitly retrieve the most relevant state-action pairs from the training videos and revises the generated procedures. To tackle high annotation cost, RAP utilizes a weakly-supervised learning manner to expand the training dataset to other task-relevant, unannotated videos by generating pseudo labels for action steps. Experiments on CrossTask and COIN benchmarks show the superiority of RAP over traditional fixed-length models, establishing it as a strong baseline solution for adaptive procedure planning.

Hung-Ting Su, Chun-Tong Chao, Ya-Ching Hsu et al.

Large Language Models (LLMs) have demonstrated effectiveness not only in language tasks but also in video reasoning. This paper introduces a novel dataset, Tropes in Movies (TiM), designed as a testbed for exploring two critical yet previously overlooked video reasoning skills: (1) Abstract Perception: understanding and tokenizing abstract concepts in videos, and (2) Long-range Compositional Reasoning: planning and integrating intermediate reasoning steps for understanding long-range videos with numerous frames. Utilizing tropes from movie storytelling, TiM evaluates the reasoning capabilities of state-of-the-art LLM-based approaches. Our experiments show that current methods, including Captioner-Reasoner, Large Multimodal Model Instruction Fine-tuning, and Visual Programming, only marginally outperform a random baseline when tackling the challenges of Abstract Perception and Long-range Compositional Reasoning. To address these deficiencies, we propose Face-Enhanced Viper of Role Interactions (FEVoRI) and Context Query Reduction (ConQueR), which enhance Visual Programming by fostering role interaction awareness and progressively refining movie contexts and trope queries during reasoning processes, significantly improving performance by 15 F1 points. However, this performance still lags behind human levels (40 vs. 65 F1). Additionally, we introduce a new protocol to evaluate the necessity of Abstract Perception and Long-range Compositional Reasoning for task resolution. This is done by analyzing the code generated through Visual Programming using an Abstract Syntax Tree (AST), thereby confirming the increased complexity of TiM. The dataset and code are available at: https://ander1119.github.io/TiM

Yulei Niu, Hanwang Zhang

Question answering (QA) models are well-known to exploit data bias, e.g., the language prior in visual QA and the position bias in reading comprehension. Recent debiasing methods achieve good out-of-distribution (OOD) generalizability with a considerable sacrifice of the in-distribution (ID) performance. Therefore, they are only applicable in domains where the test distribution is known in advance. In this paper, we present a novel debiasing method called Introspective Distillation (IntroD) to make the best of both worlds for QA. Our key technical contribution is to blend the inductive bias of OOD and ID by introspecting whether a training sample fits in the factual ID world or the counterfactual OOD one. Experiments on visual QA datasets VQA v2, VQA-CP, and reading comprehension dataset SQuAD demonstrate that our proposed IntroD maintains the competitive OOD performance compared to other debiasing methods, while sacrificing little or even achieving better ID performance compared to the non-debiasing ones.

Long Chen, Yuhang Zheng, Yulei Niu et al.

Today's VQA models still tend to capture superficial linguistic correlations in the training set and fail to generalize to the test set with different QA distributions. To reduce these language biases, recent VQA works introduce an auxiliary question-only model to regularize the training of targeted VQA model, and achieve dominating performance on diagnostic benchmarks for out-of-distribution testing. However, due to complex model design, these ensemble-based methods are unable to equip themselves with two indispensable characteristics of an ideal VQA model: 1) Visual-explainable: The model should rely on the right visual regions when making decisions. 2) Question-sensitive: The model should be sensitive to the linguistic variations in questions. To this end, we propose a novel model-agnostic Counterfactual Samples Synthesizing and Training (CSST) strategy. After training with CSST, VQA models are forced to focus on all critical objects and words, which significantly improves both visual-explainable and question-sensitive abilities. Specifically, CSST is composed of two parts: Counterfactual Samples Synthesizing (CSS) and Counterfactual Samples Training (CST). CSS generates counterfactual samples by carefully masking critical objects in images or words in questions and assigning pseudo ground-truth answers. CST not only trains the VQA models with both complementary samples to predict respective ground-truth answers, but also urges the VQA models to further distinguish the original samples and superficially similar counterfactual ones. To facilitate the CST training, we propose two variants of supervised contrastive loss for VQA, and design an effective positive and negative sample selection mechanism based on CSS. Extensive experiments have shown the effectiveness of CSST. Particularly, by building on top of model LMH+SAR, we achieve record-breaking performance on all OOD benchmarks.

Sicheng Yu, Yulei Niu, Shuohang Wang et al.

Deep neural network based question answering (QA) models are neither robust nor explainable in many cases. For example, a multiple-choice QA model, tested without any input of question, is surprisingly "capable" to predict the most of correct options. In this paper, we inspect such spurious "capability" of QA models using causal inference. We find the crux is the shortcut correlation, e.g., unrobust word alignment between passage and options learned by the models. We propose a novel approach called Counterfactual Variable Control (CVC) that explicitly mitigates any shortcut correlation and preserves the comprehensive reasoning for robust QA. Specifically, we leverage multi-branch architecture that allows us to disentangle robust and shortcut correlations in the training process of QA. We then conduct two novel CVC inference methods (on trained models) to capture the effect of comprehensive reasoning as the final prediction. For evaluation, we conduct extensive experiments using two BERT backbones on both multi-choice and span-extraction QA benchmarks. The results show that our CVC achieves high robustness against a variety of adversarial attacks in QA while maintaining good interpretation ability.

Kaihua Tang, Yulei Niu, Jianqiang Huang et al.

Today's scene graph generation (SGG) task is still far from practical, mainly due to the severe training bias, e.g., collapsing diverse "human walk on / sit on / lay on beach" into "human on beach". Given such SGG, the down-stream tasks such as VQA can hardly infer better scene structures than merely a bag of objects. However, debiasing in SGG is not trivial because traditional debiasing methods cannot distinguish between the good and bad bias, e.g., good context prior (e.g., "person read book" rather than "eat") and bad long-tailed bias (e.g., "near" dominating "behind / in front of"). In this paper, we present a novel SGG framework based on causal inference but not the conventional likelihood. We first build a causal graph for SGG, and perform traditional biased training with the graph. Then, we propose to draw the counterfactual causality from the trained graph to infer the effect from the bad bias, which should be removed. In particular, we use Total Direct Effect (TDE) as the proposed final predicate score for unbiased SGG. Note that our framework is agnostic to any SGG model and thus can be widely applied in the community who seeks unbiased predictions. By using the proposed Scene Graph Diagnosis toolkit on the SGG benchmark Visual Genome and several prevailing models, we observed significant improvements over the previous state-of-the-art methods.

Yuqi Huo, Xiaoli Xu, Yao Lu et al.

Video action recognition, which is topical in computer vision and video analysis, aims to allocate a short video clip to a pre-defined category such as brushing hair or climbing stairs. Recent works focus on action recognition with deep neural networks that achieve state-of-the-art results in need of high-performance platforms. Despite the fast development of mobile computing, video action recognition on mobile devices has not been fully discussed. In this paper, we focus on the novel mobile video action recognition task, where only the computational capabilities of mobile devices are accessible. Instead of raw videos with huge storage, we choose to extract multiple modalities (including I-frames, motion vectors, and residuals) directly from compressed videos. By employing MobileNetV2 as backbone, we propose a novel Temporal Trilinear Pooling (TTP) module to fuse the multiple modalities for mobile video action recognition. In addition to motion vectors, we also provide a temporal fusion method to explicitly induce the temporal context. The efficiency test on a mobile device indicates that our model can perform mobile video action recognition at about 40FPS. The comparative results on two benchmarks show that our model outperforms existing action recognition methods in model size and time consuming, but with competitive accuracy.

Yulei Niu, Hanwang Zhang, Zhiwu Lu et al.

We focus on grounding (i.e., localizing or linking) referring expressions in images, e.g., ``largest elephant standing behind baby elephant''. This is a general yet challenging vision-language task since it does not only require the localization of objects, but also the multimodal comprehension of context -- visual attributes (e.g., ``largest'', ``baby'') and relationships (e.g., ``behind'') that help to distinguish the referent from other objects, especially those of the same category. Due to the exponential complexity involved in modeling the context associated with multiple image regions, existing work oversimplifies this task to pairwise region modeling by multiple instance learning. In this paper, we propose a variational Bayesian method, called Variational Context, to solve the problem of complex context modeling in referring expression grounding. Specifically, our framework exploits the reciprocal relation between the referent and context, i.e., either of them influences estimation of the posterior distribution of the other, and thereby the search space of context can be greatly reduced. In addition to reciprocity, our framework considers the semantic information of context, i.e., the referring expression can be reproduced based on the estimated context. We also extend the model to unsupervised setting where no annotation for the referent is available. Extensive experiments on various benchmarks show consistent improvement over state-of-the-art methods in both supervised and unsupervised settings.

Hanwang Zhang, Yulei Niu, Shih-Fu Chang

We focus on grounding (i.e., localizing or linking) referring expressions in images, e.g., "largest elephant standing behind baby elephant". This is a general yet challenging vision-language task since it does not only require the localization of objects, but also the multimodal comprehension of context --- visual attributes (e.g., "largest", "baby") and relationships (e.g., "behind") that help to distinguish the referent from other objects, especially those of the same category. Due to the exponential complexity involved in modeling the context associated with multiple image regions, existing work oversimplifies this task to pairwise region modeling by multiple instance learning. In this paper, we propose a variational Bayesian method, called Variational Context, to solve the problem of complex context modeling in referring expression grounding. Our model exploits the reciprocal relation between the referent and context, i.e., either of them influences the estimation of the posterior distribution of the other, and thereby the search space of context can be greatly reduced, resulting in better localization of referent. We develop a novel cue-specific language-vision embedding network that learns this reciprocity model end-to-end. We also extend the model to the unsupervised setting where no annotation for the referent is available. Extensive experiments on various benchmarks show consistent improvement over state-of-the-art methods in both supervised and unsupervised settings.

Yulei Niu, Zhiwu Lu, Ji-Rong Wen et al.

Image annotation aims to annotate a given image with a variable number of class labels corresponding to diverse visual concepts. In this paper, we address two main issues in large-scale image annotation: 1) how to learn a rich feature representation suitable for predicting a diverse set of visual concepts ranging from object, scene to abstract concept; 2) how to annotate an image with the optimal number of class labels. To address the first issue, we propose a novel multi-scale deep model for extracting rich and discriminative features capable of representing a wide range of visual concepts. Specifically, a novel two-branch deep neural network architecture is proposed which comprises a very deep main network branch and a companion feature fusion network branch designed for fusing the multi-scale features computed from the main branch. The deep model is also made multi-modal by taking noisy user-provided tags as model input to complement the image input. For tackling the second issue, we introduce a label quantity prediction auxiliary task to the main label prediction task to explicitly estimate the optimal label number for a given image. Extensive experiments are carried out on two large-scale image annotation benchmark datasets and the results show that our method significantly outperforms the state-of-the-art.