Yuta Nakashima

Dianxing Shi, Junqi He, Junhao Chen et al.

Self-evolving agents improve through continual self-play and self-generated learning signals, but autonomous evolution can also cause capability degradation and safety drift. Although human feedback has proven effective for static and post-trained agents, its role in self-evolving systems remains underexplored. We introduce Agent Norm Correction through Human-like Oversight and Review (ANCHOR), an LLM-based framework that simulates human supervision and delivers feedback at various phases of self-evolution. With ANCHOR, we evaluate two representative open-source self-evolving agent systems across coding, mathematical reasoning, and safety. Our results show that even limited supervision substantially mitigates safety degradation while preserving stable performance on core evolutionary objectives. Further analysis shows that supervision over the output verification phase is the most effective for intervention, whereas increasing supervision frequency yields diminishing returns. These findings provide empirical evidence and practical guidance for designing more stable, controllable, and human-aligned self-evolving agent systems.

Bowen Wang, Liangzhi Li, Yuta Nakashima et al.

Interpreting and explaining the behavior of deep neural networks is critical for many tasks. Explainable AI provides a way to address this challenge, mostly by providing per-pixel relevance to the decision. Yet, interpreting such explanations may require expert knowledge. Some recent attempts toward interpretability adopt a concept-based framework, giving a higher-level relationship between some concepts and model decisions. This paper proposes Bottleneck Concept Learner (BotCL), which represents an image solely by the presence/absence of concepts learned through training over the target task without explicit supervision over the concepts. It uses self-supervision and tailored regularizers so that learned concepts can be human-understandable. Using some image classification tasks as our testbed, we demonstrate BotCL's potential to rebuild neural networks for better interpretability. Code is available at https://github.com/wbw520/BotCL and a simple demo is available at https://botcl.liangzhili.com/.

Zongshang Pang, Yuta Nakashima, Mayu Otani et al.

Video summarization aims to select the most informative subset of frames in a video to facilitate efficient video browsing. Unsupervised methods usually rely on heuristic training objectives such as diversity and representativeness. However, such methods need to bootstrap the online-generated summaries to compute the objectives for importance score regression. We consider such a pipeline inefficient and seek to directly quantify the frame-level importance with the help of contrastive losses in the representation learning literature. Leveraging the contrastive losses, we propose three metrics featuring a desirable key frame: local dissimilarity, global consistency, and uniqueness. With features pre-trained on the image classification task, the metrics can already yield high-quality importance scores, demonstrating competitive or better performance than past heavily-trained methods. We show that by refining the pre-trained features with a lightweight contrastively learned projection module, the frame-level importance scores can be further improved, and the model can also leverage a large number of random videos and generalize to test videos with decent performance. Code available at https://github.com/pangzss/pytorch-CTVSUM.

Yusuke Hirota, Yuta Nakashima, Noa Garcia

Image captioning models are known to perpetuate and amplify harmful societal bias in the training set. In this work, we aim to mitigate such gender bias in image captioning models. While prior work has addressed this problem by forcing models to focus on people to reduce gender misclassification, it conversely generates gender-stereotypical words at the expense of predicting the correct gender. From this observation, we hypothesize that there are two types of gender bias affecting image captioning models: 1) bias that exploits context to predict gender, and 2) bias in the probability of generating certain (often stereotypical) words because of gender. To mitigate both types of gender biases, we propose a framework, called LIBRA, that learns from synthetically biased samples to decrease both types of biases, correcting gender misclassification and changing gender-stereotypical words to more neutral ones. Code is available at https://github.com/rebnej/LIBRA.

Jiahao Zhang, Bowen Wang, Liangzhi Li et al.

Large-scale models trained on extensive datasets, have emerged as the preferred approach due to their high generalizability across various tasks. In-context learning (ICL), a popular strategy in natural language processing, uses such models for different tasks by providing instructive prompts but without updating model parameters. This idea is now being explored in computer vision, where an input-output image pair (called an in-context pair) is supplied to the model with a query image as a prompt to exemplify the desired output. The efficacy of visual ICL often depends on the quality of the prompts. We thus introduce a method coined Instruct Me More (InMeMo), which augments in-context pairs with a learnable perturbation (prompt), to explore its potential. Our experiments on mainstream tasks reveal that InMeMo surpasses the current state-of-the-art performance. Specifically, compared to the baseline without learnable prompt, InMeMo boosts mIoU scores by 7.35 and 15.13 for foreground segmentation and single object detection tasks, respectively. Our findings suggest that InMeMo offers a versatile and efficient way to enhance the performance of visual ICL with lightweight training. Code is available at https://github.com/Jackieam/InMeMo.

Bowen Wang, Jiaxing Zhang, Ran Zhang et al.

Facade parsing stands as a pivotal computer vision task with far-reaching applications in areas like architecture, urban planning, and energy efficiency. Despite the recent success of deep learning-based methods in yielding impressive results on certain open-source datasets, their viability for real-world applications remains uncertain. Real-world scenarios are considerably more intricate, demanding greater computational efficiency. Existing datasets often fall short in representing these settings, and previous methods frequently rely on extra models to enhance accuracy, which requires much computation cost. In this paper, we introduce Comprehensive Facade Parsing (CFP), a dataset meticulously designed to encompass the intricacies of real-world facade parsing tasks. Comprising a total of 602 high-resolution street-view images, this dataset captures a diverse array of challenging scenarios, including sloping angles and densely clustered buildings, with painstakingly curated annotations for each image. We introduce a new pipeline known as Revision-based Transformer Facade Parsing (RTFP). This marks the pioneering utilization of Vision Transformers (ViT) in facade parsing, and our experimental results definitively substantiate its merit. We also design Line Acquisition, Filtering, and Revision (LAFR), an efficient yet accurate revision algorithm that can improve the segment result solely from simple line detection using prior knowledge of the facade. In ECP 2011, RueMonge 2014, and our CFP, we evaluate the superiority of our method.

Mayu Otani, Riku Togashi, Yu Sawai et al.

Human evaluation is critical for validating the performance of text-to-image generative models, as this highly cognitive process requires deep comprehension of text and images. However, our survey of 37 recent papers reveals that many works rely solely on automatic measures (e.g., FID) or perform poorly described human evaluations that are not reliable or repeatable. This paper proposes a standardized and well-defined human evaluation protocol to facilitate verifiable and reproducible human evaluation in future works. In our pilot data collection, we experimentally show that the current automatic measures are incompatible with human perception in evaluating the performance of the text-to-image generation results. Furthermore, we provide insights for designing human evaluation experiments reliably and conclusively. Finally, we make several resources publicly available to the community to facilitate easy and fast implementations.

Mayu Otani, Riku Togashi, Yuta Nakashima et al.

Mean Average Precision (mAP) is the primary evaluation measure for object detection. Although object detection has a broad range of applications, mAP evaluates detectors in terms of the performance of ranked instance retrieval. Such the assumption for the evaluation task does not suit some downstream tasks. To alleviate the gap between downstream tasks and the evaluation scenario, we propose Optimal Correction Cost (OC-cost), which assesses detection accuracy at image level. OC-cost computes the cost of correcting detections to ground truths as a measure of accuracy. The cost is obtained by solving an optimal transportation problem between the detections and the ground truths. Unlike mAP, OC-cost is designed to penalize false positive and false negative detections properly, and every image in a dataset is treated equally. Our experimental result validates that OC-cost has better agreement with human preference than a ranking-based measure, i.e., mAP for a single image. We also show that detectors' rankings by OC-cost are more consistent on different data splits than mAP. Our goal is not to replace mAP with OC-cost but provide an additional tool to evaluate detectors from another aspect. To help future researchers and developers choose a target measure, we provide a series of experiments to clarify how mAP and OC-cost differ.

Junhao Chen, Bowen Wang, Zhouqiang Jiang et al.

Large Language Models (LLMs) have demonstrated significant capabilities, particularly in the domain of question answering (QA). However, their effectiveness in QA is often undermined by the vagueness of user questions. To address this issue, we introduce single-round instance-level prompt optimization, referred to as question rewriter. By enhancing the intelligibility of human questions for black-box LLMs, our question rewriter improves the quality of generated answers. The rewriter is optimized using direct preference optimization based on feedback collected from automatic criteria for evaluating generated answers; therefore, its training does not require costly human annotations. The experiments across multiple black-box LLMs and long-form question answering (LFQA) datasets demonstrate the efficacy of our method. This paper provides a practical framework for training question rewriters and sets a precedent for future explorations in prompt optimization within LFQA tasks. Code is available at https://github.com/3244we/Question-Rewriter.

Noa Garcia, Yusuke Hirota, Yankun Wu et al.

The increasing tendency to collect large and uncurated datasets to train vision-and-language models has raised concerns about fair representations. It is known that even small but manually annotated datasets, such as MSCOCO, are affected by societal bias. This problem, far from being solved, may be getting worse with data crawled from the Internet without much control. In addition, the lack of tools to analyze societal bias in big collections of images makes addressing the problem extremely challenging. Our first contribution is to annotate part of the Google Conceptual Captions dataset, widely used for training vision-and-language models, with four demographic and two contextual attributes. Our second contribution is to conduct a comprehensive analysis of the annotations, focusing on how different demographic groups are represented. Our last contribution lies in evaluating three prevailing vision-and-language tasks: image captioning, text-image CLIP embeddings, and text-to-image generation, showing that societal bias is a persistent problem in all of them.

Yusuke Hirota, Yuta Nakashima, Noa Garcia

Vision-and-language tasks have increasingly drawn more attention as a means to evaluate human-like reasoning in machine learning models. A popular task in the field is visual question answering (VQA), which aims to answer questions about images. However, VQA models have been shown to exploit language bias by learning the statistical correlations between questions and answers without looking into the image content: e.g., questions about the color of a banana are answered with yellow, even if the banana in the image is green. If societal bias (e.g., sexism, racism, ableism, etc.) is present in the training data, this problem may be causing VQA models to learn harmful stereotypes. For this reason, we investigate gender and racial bias in five VQA datasets. In our analysis, we find that the distribution of answers is highly different between questions about women and men, as well as the existence of detrimental gender-stereotypical samples. Likewise, we identify that specific race-related attributes are underrepresented, whereas potentially discriminatory samples appear in the analyzed datasets. Our findings suggest that there are dangers associated to using VQA datasets without considering and dealing with the potentially harmful stereotypes. We conclude the paper by proposing solutions to alleviate the problem before, during, and after the dataset collection process.

Yusuke Hirota, Yuta Nakashima, Noa Garcia

We study societal bias amplification in image captioning. Image captioning models have been shown to perpetuate gender and racial biases, however, metrics to measure, quantify, and evaluate the societal bias in captions are not yet standardized. We provide a comprehensive study on the strengths and limitations of each metric, and propose LIC, a metric to study captioning bias amplification. We argue that, for image captioning, it is not enough to focus on the correct prediction of the protected attribute, and the whole context should be taken into account. We conduct extensive evaluation on traditional and state-of-the-art image captioning models, and surprisingly find that, by only focusing on the protected attribute prediction, bias mitigation models are unexpectedly amplifying bias.

Yankun Wu, Yuta Nakashima, Noa Garcia

The duality of content and style is inherent to the nature of art. For humans, these two elements are clearly different: content refers to the objects and concepts in the piece of art, and style to the way it is expressed. This duality poses an important challenge for computer vision. The visual appearance of objects and concepts is modulated by the style that may reflect the author's emotions, social trends, artistic movement, etc., and their deep comprehension undoubtfully requires to handle both. A promising step towards a general paradigm for art analysis is to disentangle content and style, whereas relying on human annotations to cull a single aspect of artworks has limitations in learning semantic concepts and the visual appearance of paintings. We thus present GOYA, a method that distills the artistic knowledge captured in a recent generative model to disentangle content and style. Experiments show that synthetically generated images sufficiently serve as a proxy of the real distribution of artworks, allowing GOYA to separately represent the two elements of art while keeping more information than existing methods.

Bowen Wang, Jiuyang Chang, Yiming Qian et al.

Large language models (LLMs) have recently showcased remarkable capabilities, spanning a wide range of tasks and applications, including those in the medical domain. Models like GPT-4 excel in medical question answering but may face challenges in the lack of interpretability when handling complex tasks in real clinical settings. We thus introduce the diagnostic reasoning dataset for clinical notes (DiReCT), aiming at evaluating the reasoning ability and interpretability of LLMs compared to human doctors. It contains 511 clinical notes, each meticulously annotated by physicians, detailing the diagnostic reasoning process from observations in a clinical note to the final diagnosis. Additionally, a diagnostic knowledge graph is provided to offer essential knowledge for reasoning, which may not be covered in the training data of existing LLMs. Evaluations of leading LLMs on DiReCT bring out a significant gap between their reasoning ability and that of human doctors, highlighting the critical need for models that can reason effectively in real-world clinical scenarios.

Yusuke Hirota, Jerone T. A. Andrews, Dora Zhao et al.

We tackle societal bias in image-text datasets by removing spurious correlations between protected groups and image attributes. Traditional methods only target labeled attributes, ignoring biases from unlabeled ones. Using text-guided inpainting models, our approach ensures protected group independence from all attributes and mitigates inpainting biases through data filtering. Evaluations on multi-label image classification and image captioning tasks show our method effectively reduces bias without compromising performance across various models.

Tianwei Chen, Noa Garcia, Mayu Otani et al.

Is more data always better to train vision-and-language models? We study knowledge transferability in multi-modal tasks. The current tendency in machine learning is to assume that by joining multiple datasets from different tasks their overall performance will improve. However, we show that not all the knowledge transfers well or has a positive impact on related tasks, even when they share a common goal. We conduct an exhaustive analysis based on hundreds of cross-experiments on 12 vision-and-language tasks categorized in 4 groups. Whereas tasks in the same group are prone to improve each other, results show that this is not always the case. Other factors such as dataset size or pre-training stage have also a great impact on how well the knowledge is transferred.

Hugo Lemarchant, Liangzi Li, Yiming Qian et al.

Vision Transformers (ViTs) are becoming a very popular paradigm for vision tasks as they achieve state-of-the-art performance on image classification. However, although early works implied that this network structure had increased robustness against adversarial attacks, some works argue ViTs are still vulnerable. This paper presents our first attempt toward detecting adversarial attacks during inference time using the network's input and outputs as well as latent features. We design four quantifications (or derivatives) of input, output, and latent vectors of ViT-based models that provide a signature of the inference, which could be beneficial for the attack detection, and empirically study their behavior over clean samples and adversarial samples. The results demonstrate that the quantifications from input (images) and output (posterior probabilities) are promising for distinguishing clean and adversarial samples, while latent vectors offer less discriminative power, though they give some insights on how adversarial perturbations work.

Lu Wei, Yuta Nakashima, Noa Garcia

The widespread adoption of text-to-image (T2I) generation has raised concerns about privacy, bias, and copyright violations. Concept erasure techniques offer a promising solution by selectively removing undesired concepts from pre-trained models without requiring full retraining. However, these methods are often evaluated on a limited set of concepts, relying on overly simplistic and direct prompts. To test the boundaries of concept erasure techniques, and assess whether they truly remove targeted concepts from model representations, we introduce EMMA, a benchmark that evaluates five key dimensions of concept erasure over 12 metrics. EMMA goes beyond standard metrics like image quality and time efficiency, testing robustness under challenging conditions, including indirect descriptions, visually similar non-target concepts, and potential gender and ethnicity bias, providing a socially aware analysis of method behavior. Using EMMA, we analyze five concept erasure methods across five domains (objects, celebrities, art styles, NSFW, and copyright). Our results show that existing methods struggle with implicit prompts (i.e., generating the erased concept when it is indirectly referenced) and visually similar non-target concepts (i.e., failing to generate non-targeted concepts resembling the erased one), while some amplify gender and ethnicity bias compared to the original model.

Bowen Wang, Liangzhi Li, Jiahao Zhang et al.

The imperative to comprehend the behaviors of deep learning models is of utmost importance. In this realm, Explainable Artificial Intelligence (XAI) has emerged as a promising avenue, garnering increasing interest in recent years. Despite this, most existing methods primarily depend on gradients or input perturbation, which often fails to embed explanations directly within the model's decision-making process. Addressing this gap, we introduce ESCOUTER, a visually explainable classifier based on the modified slot attention mechanism. ESCOUTER distinguishes itself by not only delivering high classification accuracy but also offering more transparent insights into the reasoning behind its decisions. It differs from prior approaches in two significant aspects: (a) ESCOUTER incorporates explanations into the final confidence scores for each category, providing a more intuitive interpretation, and (b) it offers positive or negative explanations for all categories, elucidating "why an image belongs to a certain category" or "why it does not." A novel loss function specifically for ESCOUTER is designed to fine-tune the model's behavior, enabling it to toggle between positive and negative explanations. Moreover, an area loss is also designed to adjust the size of the explanatory regions for a more precise explanation. Our method, rigorously tested across various datasets and XAI metrics, outperformed previous state-of-the-art methods, solidifying its effectiveness as an explanatory tool.

Yusuke Hirota, Min-Hung Chen, Chien-Yi Wang et al.

Large-scale vision-language models, such as CLIP, are known to contain societal bias regarding protected attributes (e.g., gender, age). This paper aims to address the problems of societal bias in CLIP. Although previous studies have proposed to debias societal bias through adversarial learning or test-time projecting, our comprehensive study of these works identifies two critical limitations: 1) loss of attribute information when it is explicitly disclosed in the input and 2) use of the attribute annotations during debiasing process. To mitigate societal bias in CLIP and overcome these limitations simultaneously, we introduce a simple-yet-effective debiasing method called SANER (societal attribute neutralizer) that eliminates attribute information from CLIP text features only of attribute-neutral descriptions. Experimental results show that SANER, which does not require attribute annotations and preserves original information for attribute-specific descriptions, demonstrates superior debiasing ability than the existing methods.

Jiahao Zhang, Bowen Wang, Hong Liu et al.

Large-scale models trained on extensive datasets have become the standard due to their strong generalizability across diverse tasks. In-context learning (ICL), widely used in natural language processing, leverages these models by providing task-specific prompts without modifying their parameters. This paradigm is increasingly being adapted for computer vision, where models receive an input-output image pair, known as an in-context pair, alongside a query image to illustrate the desired output. However, the success of visual ICL largely hinges on the quality of these prompts. To address this, we propose Enhanced Instruct Me More (E-InMeMo), a novel approach that incorporates learnable perturbations into in-context pairs to optimize prompting. Through extensive experiments on standard vision tasks, E-InMeMo demonstrates superior performance over existing state-of-the-art methods. Notably, it improves mIoU scores by 7.99 for foreground segmentation and by 17.04 for single object detection when compared to the baseline without learnable prompts. These results highlight E-InMeMo as a lightweight yet effective strategy for enhancing visual ICL. Code is publicly available at: https://github.com/Jackieam/E-InMeMo

Rawisara Lohanimit, Yankun Wu, Amelia Katirai et al.

The rise of generative models has led to increased use of large-scale datasets collected from the internet, often with minimal or no data curation. This raises concerns about the inclusion of sensitive or private information. In this work, we explore the presence of pregnancy ultrasound images, which contain sensitive personal information and are often shared online. Through a systematic examination of LAION-400M dataset using CLIP embedding similarity, we retrieve images containing pregnancy ultrasound and detect thousands of entities of private information such as names and locations. Our findings reveal that multiple images have high-risk information that could enable re-identification or impersonation. We conclude with recommended practices for dataset curation, data privacy, and ethical use of public image datasets.

Ryan Ramos, Vladan Stojnić, Giorgos Kordopatis-Zilos et al.

Prior work has analyzed the robustness of visual encoders to image transformations and corruptions, particularly in cases where such alterations are not seen during training. When this occurs, they introduce a form of distribution shift at test time, often leading to performance degradation. The primary focus has been on severe corruptions that, when applied aggressively, distort useful signals necessary for accurate semantic predictions. We take a different perspective by analyzing parameters of the image acquisition process and transformations that may be subtle or even imperceptible to the human eye. We find that such parameters are systematically encoded in the learned visual representations and can be easily recovered. More strikingly, their presence can have a profound impact, either positively or negatively, on semantic predictions. This effect depends on whether there is a strong correlation or anti-correlation between semantic labels and these acquisition-based or processing-based labels. Our code and data are available at: https://github.com/ryan-caesar-ramos/visual-encoder-traces

Patrick Ramos, Nicolas Gonthier, Selina Khan et al.

Object detection in art is a valuable tool for the digital humanities, as it allows for faster identification of objects in artistic and historical images compared to humans. However, annotating such images poses significant challenges due to the need for specialized domain expertise. We present NADA (no annotations for detection in art), a pipeline that leverages diffusion models' art-related knowledge for object detection in paintings without the need for full bounding box supervision. Our method, which supports both weakly-supervised and zero-shot scenarios and does not require any fine-tuning of its pretrained components, consists of a class proposer based on large vision-language models and a class-conditioned detector based on Stable Diffusion. NADA is evaluated on two artwork datasets, ArtDL 2.0 and IconArt, outperforming prior work in weakly-supervised detection, while being the first work for zero-shot object detection in art. Code is available at https://github.com/patrick-john-ramos/nada

Liangzhi Li, Manisha Verma, Bowen Wang et al.

The status of retinal arteriovenous crossing is of great significance for clinical evaluation of arteriolosclerosis and systemic hypertension. As an ophthalmology diagnostic criteria, Scheie's classification has been used to grade the severity of arteriolosclerosis. In this paper, we propose a deep learning approach to support the diagnosis process, which, to the best of our knowledge, is one of the earliest attempts in medical imaging. The proposed pipeline is three-fold. First, we adopt segmentation and classification models to automatically obtain vessels in a retinal image with the corresponding artery/vein labels and find candidate arteriovenous crossing points. Second, we use a classification model to validate the true crossing point. At last, the grade of severity for the vessel crossings is classified. To better address the problem of label ambiguity and imbalanced label distribution, we propose a new model, named multi-diagnosis team network (MDTNet), in which the sub-models with different structures or different loss functions provide different decisions. MDTNet unifies these diverse theories to give the final decision with high accuracy. Our severity grading method was able to validate crossing points with precision and recall of 96.3% and 96.3%, respectively. Among correctly detected crossing points, the kappa value for the agreement between the grading by a retina specialist and the estimated score was 0.85, with an accuracy of 0.92. The numerical results demonstrate that our method can achieve a good performance in both arteriovenous crossing validation and severity grading tasks. By the proposed models, we could build a pipeline reproducing retina specialist's subjective grading without feature extractions. The code is available for reproducibility.

Chenhui Chu, Yuto Takebayashi, Mishra Vipul et al.

A visual relationship denotes a relationship between two objects in an image, which can be represented as a triplet of (subject; predicate; object). Visual relationship detection is crucial for scene understanding in images. Existing visual relationship detection datasets only contain true relationships that correctly describe the content in an image. However, distinguishing false visual relationships from true ones is also crucial for image understanding and grounded natural language processing. In this paper, we construct a visual relationship authenticity dataset, where both true and false relationships among all objects appeared in the captions in the Flickr30k entities image caption dataset are annotated. The dataset is available at https://github.com/codecreator2053/VR_ClassifiedDataset. We hope that this dataset can promote the study on both vision and language understanding.

Liangzhi Li, Manisha Verma, Yuta Nakashima et al.

Retinal vessel segmentation is of great interest for diagnosis of retinal vascular diseases. To further improve the performance of vessel segmentation, we propose IterNet, a new model based on UNet, with the ability to find obscured details of the vessel from the segmented vessel image itself, rather than the raw input image. IterNet consists of multiple iterations of a mini-UNet, which can be 4$\times$ deeper than the common UNet. IterNet also adopts the weight-sharing and skip-connection features to facilitate training; therefore, even with such a large architecture, IterNet can still learn from merely 10$\sim$20 labeled images, without pre-training or any prior knowledge. IterNet achieves AUCs of 0.9816, 0.9851, and 0.9881 on three mainstream datasets, namely DRIVE, CHASE-DB1, and STARE, respectively, which currently are the best scores in the literature. The source code is available.

Yankun Wu, Yuta Nakashima, Noa Garcia

Several studies have raised awareness about social biases in image generative models, demonstrating their predisposition towards stereotypes and imbalances. This paper contributes to this growing body of research by introducing an evaluation protocol that analyzes the impact of gender indicators at every step of the generation process on Stable Diffusion images. Leveraging insights from prior work, we explore how gender indicators not only affect gender presentation but also the representation of objects and layouts within the generated images. Our findings include the existence of differences in the depiction of objects, such as instruments tailored for specific genders, and shifts in overall layouts. We also reveal that neutral prompts tend to produce images more aligned with masculine prompts than their feminine counterparts. We further explore where bias originates through representational disparities and how it manifests in the images via prompt-image dependencies, and provide recommendations for developers and users to mitigate potential bias in image generation.

Tianwei Chen, Yusuke Hirota, Mayu Otani et al.

We investigate the impact of deep generative models on potential social biases in upcoming computer vision models. As the internet witnesses an increasing influx of AI-generated images, concerns arise regarding inherent biases that may accompany them, potentially leading to the dissemination of harmful content. This paper explores whether a detrimental feedback loop, resulting in bias amplification, would occur if generated images were used as the training data for future models. We conduct simulations by progressively substituting original images in COCO and CC3M datasets with images generated through Stable Diffusion. The modified datasets are used to train OpenCLIP and image captioning models, which we evaluate in terms of quality and bias. Contrary to expectations, our findings indicate that introducing generated images during training does not uniformly amplify bias. Instead, instances of bias mitigation across specific tasks are observed. We further explore the factors that may influence these phenomena, such as artifacts in image generation (e.g., blurry faces) or pre-existing biases in the original datasets.

Jiahao Zhang, Ryota Yoshihashi, Shunsuke Kitada et al.

Large language models (LLMs) have proven effective for layout generation due to their ability to produce structure-description languages, such as HTML or JSON. In this paper, we argue that while LLMs can perform reasonably well in certain cases, their intrinsic limitation of not being able to perceive images restricts their effectiveness in tasks requiring visual content, e.g., content-aware layout generation. Therefore, we explore whether large vision-language models (LVLMs) can be applied to content-aware layout generation. To this end, inspired by the iterative revision and heuristic evaluation workflow of designers, we propose the training-free Visual-Aware Self-Correction LAyout GeneRation (VASCAR). VASCAR enables LVLMs (e.g., GPT-4o and Gemini) iteratively refine their outputs with reference to rendered layout images, which are visualized as colored bounding boxes on poster background (i.e., canvas). Extensive experiments and user study demonstrate VASCAR's effectiveness, achieving state-of-the-art (SOTA) layout generation quality. Furthermore, the generalizability of VASCAR across GPT-4o and Gemini demonstrates its versatility.

Zhouqiang Jiang, Bowen Wang, Junhao Chen et al.

Recent approaches for visually-rich document understanding (VrDU) uses manually annotated semantic groups, where a semantic group encompasses all semantically relevant but not obviously grouped words. As OCR tools are unable to automatically identify such grouping, we argue that current VrDU approaches are unrealistic. We thus introduce a new variant of the VrDU task, real-world visually-rich document understanding (ReVrDU), that does not allow for using manually annotated semantic groups. We also propose a new method, ReLayout, compliant with the ReVrDU scenario, which learns to capture semantic grouping through arranging words and bringing the representations of words that belong to the potential same semantic group closer together. Our experimental results demonstrate the performance of existing methods is deteriorated with the ReVrDU task, while ReLayout shows superiour performance.

Zongshang Pang, Mayu Otani, Yuta Nakashima

Localizing user-queried events through natural language is crucial for video understanding models. Recent methods predominantly adapt Video LLMs to generate event boundary timestamps to handle temporal localization tasks, which struggle to leverage LLMs' powerful semantic understanding. In this work, we introduce MeCo, a novel timestamp-free framework that enables video LLMs to fully harness their intrinsic semantic capabilities for temporal localization tasks. Rather than outputting boundary timestamps, MeCo partitions videos into holistic event and transition segments based on the proposed structural token generation and grounding pipeline, derived from video LLMs' temporal structure understanding capability. We further propose a query-focused captioning task that compels the LLM to extract fine-grained, event-specific details, bridging the gap between localization and higher-level semantics and enhancing localization performance. Extensive experiments on diverse temporal localization tasks show that MeCo consistently outperforms boundary-centric methods, underscoring the benefits of a semantic-driven approach for temporal localization with video LLMs.

Jiahao Zhang, Bowen Wang, Hong Liu et al.

Visual In-Context Learning (VICL) uses input-output image pairs, referred to as in-context pairs (or examples), as prompts alongside query images to guide models in performing diverse vision tasks. However, VICL often suffers from over-reliance on a single in-context pair, which can lead to biased and unstable predictions. We introduce PAtch-based $k$-Nearest neighbor visual In-Context Learning (PANICL), a general training-free framework that mitigates this issue by leveraging multiple in-context pairs. PANICL smooths assignment scores across pairs, reducing bias without requiring additional training. Extensive experiments on a variety of tasks, including foreground segmentation, single object detection, colorization, multi-object segmentation, and keypoint detection, demonstrate consistent improvements over strong baselines. Moreover, PANICL exhibits strong robustness to domain shifts, including dataset-level shift (e.g., from COCO to Pascal) and label-space shift (e.g., FSS-1000), and generalizes well to other VICL models such as SegGPT, Painter, and LVM, highlighting its versatility and broad applicability.

Bowen Wang, Zhouqiang Jiang, Yasuaki Susumu et al.

The real value of knowledge lies not just in its accumulation, but in its potential to be harnessed effectively to conquer the unknown. Although recent multimodal large language models (MLLMs) exhibit impressing multimodal capabilities, they often fail in rarely encountered domain-specific tasks due to limited relevant knowledge. To explore this, we adopt visual game cognition as a testbed and select Monster Hunter: World as the target to construct a multimodal knowledge graph (MH-MMKG), which incorporates multi-modalities and intricate entity relations. We also design a series of challenging queries based on MH-MMKG to evaluate the models' ability for complex knowledge retrieval and reasoning. Furthermore, we propose a multi-agent retriever that enables a model to autonomously search relevant knowledge without additional training. Experimental results show that our approach significantly enhances the performance of MLLMs, providing a new perspective on multimodal knowledge-augmented reasoning and laying a solid foundation for future research.

Yusuke Hirota, Ryo Hachiuma, Boyi Li et al. · uw

Gender bias in vision-language foundation models (VLMs) raises concerns about their safe deployment and is typically evaluated using benchmarks with gender annotations on real-world images. However, as these benchmarks often contain spurious correlations between gender and non-gender features, such as objects and backgrounds, we identify a critical oversight in gender bias evaluation: Do spurious features distort gender bias evaluation? To address this question, we systematically perturb non-gender features across four widely used benchmarks (COCO-gender, FACET, MIAP, and PHASE) and various VLMs to quantify their impact on bias evaluation. Our findings reveal that even minimal perturbations, such as masking just 10% of objects or weakly blurring backgrounds, can dramatically alter bias scores, shifting metrics by up to 175% in generative VLMs and 43% in CLIP variants. This suggests that current bias evaluations often reflect model responses to spurious features rather than gender bias, undermining their reliability. Since creating spurious feature-free benchmarks is fundamentally challenging, we recommend reporting bias metrics alongside feature-sensitivity measurements to enable a more reliable bias assessment.

Ryan Ramos, Yusuke Hirota, Yuta Nakashima et al.

The recycling of contrastive language-image pre-trained (CLIP) models as backbones for a large number of downstream tasks calls for a thorough analysis of their transferability implications, especially their well-documented reproduction of social biases and human stereotypes. How do such biases, learned during pre-training, propagate to downstream applications like visual question answering or image captioning? Do they transfer at all? We investigate this phenomenon, referred to as bias transfer in prior literature, through a comprehensive empirical analysis. Firstly, we examine how pre-training bias varies between global and local views of data, finding that bias measurement is highly dependent on the subset of data on which it is computed. Secondly, we analyze correlations between biases in the pre-trained models and the downstream tasks across varying levels of pre-training bias, finding difficulty in discovering consistent trends in bias transfer. Finally, we explore why this inconsistency occurs, showing that under the current paradigm, representation spaces of different pre-trained CLIPs tend to converge when adapted for downstream tasks. We hope this work offers valuable insights into bias behavior and informs future research to promote better bias mitigation practices.

Yusuke Hirota, Boyi Li, Ryo Hachiuma et al.

Large Vision-Language Models (LVLMs) have transformed image captioning, shifting from concise captions to detailed descriptions. We introduce LOTUS, a leaderboard for evaluating detailed captions, addressing three main gaps in existing evaluations: lack of standardized criteria, bias-aware assessments, and user preference considerations. LOTUS comprehensively evaluates various aspects, including caption quality (e.g., alignment, descriptiveness), risks (\eg, hallucination), and societal biases (e.g., gender bias) while enabling preference-oriented evaluations by tailoring criteria to diverse user preferences. Our analysis of recent LVLMs reveals no single model excels across all criteria, while correlations emerge between caption detail and bias risks. Preference-oriented evaluations demonstrate that optimal model selection depends on user priorities.

Yusuke Hirota, Ryo Hachiuma, Chao-Han Huck Yang et al.

Large language models (LLMs) have enhanced the capacity of vision-language models to caption visual text. This generative approach to image caption enrichment further makes textual captions more descriptive, improving alignment with the visual context. However, while many studies focus on benefits of generative caption enrichment (GCE), are there any negative side effects? We compare standard-format captions and recent GCE processes from the perspectives of "gender bias" and "hallucination", showing that enriched captions suffer from increased gender bias and hallucination. Furthermore, models trained on these enriched captions amplify gender bias by an average of 30.9% and increase hallucination by 59.5%. This study serves as a caution against the trend of making captions more descriptive.

Wanqing Zhao, Yuta Nakashima, Haiyuan Chen et al.

Fake news detection in social media has become increasingly important due to the rapid proliferation of personal media channels and the consequential dissemination of misleading information. Existing methods, which primarily rely on multimodal features and graph-based techniques, have shown promising performance in detecting fake news. However, they still face a limitation, i.e., sparsity in graph connections, which hinders capturing possible interactions among tweets. This challenge has motivated us to explore a novel method that densifies the graph's connectivity to capture denser interaction better. Our method constructs a cross-modal tweet graph using CLIP, which encodes images and text into a unified space, allowing us to extract potential connections based on similarities in text and images. We then design a Feature Contextualization Network with Label Propagation (FCN-LP) to model the interaction among tweets as well as positive or negative correlations between predicted labels of connected tweets. The propagated labels from the graph are weighted and aggregated for the final detection. To enhance the model's generalization ability to unseen events, we introduce a domain generalization loss that ensures consistent features between tweets on seen and unseen events. We use three publicly available fake news datasets, Twitter, PHEME, and Weibo, for evaluation. Our method consistently improves the performance over the state-of-the-art methods on all benchmark datasets and effectively demonstrates its aptitude for generalizing fake news detection in social media.

Riku Togashi, Mayu Otani, Yuta Nakashima et al.

Evaluation measures have a crucial impact on the direction of research. Therefore, it is of utmost importance to develop appropriate and reliable evaluation measures for new applications where conventional measures are not well suited. Video Moment Retrieval (VMR) is one such application, and the current practice is to use R@$K,θ$ for evaluating VMR systems. However, this measure has two disadvantages. First, it is rank-insensitive: It ignores the rank positions of successfully localised moments in the top-$K$ ranked list by treating the list as a set. Second, it binarizes the Intersection over Union (IoU) of each retrieved video moment using the threshold $θ$ and thereby ignoring fine-grained localisation quality of ranked moments. We propose an alternative measure for evaluating VMR, called Average Max IoU (AxIoU), which is free from the above two problems. We show that AxIoU satisfies two important axioms for VMR evaluation, namely, \textbf{Invariance against Redundant Moments} and \textbf{Monotonicity with respect to the Best Moment}, and also that R@$K,θ$ satisfies the first axiom only. We also empirically examine how AxIoU agrees with R@$K,θ$, as well as its stability with respect to change in the test data and human-annotated temporal boundaries.

Tianran Wu, Noa Garcia, Mayu Otani et al.

Video question answering (VideoQA) is designed to answer a given question based on a relevant video clip. The current available large-scale datasets have made it possible to formulate VideoQA as the joint understanding of visual and language information. However, this training procedure is costly and still less competent with human performance. In this paper, we investigate a transfer learning method by the introduction of domain-agnostic knowledge and domain-specific knowledge. First, we develop a novel transfer learning framework, which finetunes the pre-trained model by applying domain-agnostic knowledge as the medium. Second, we construct a new VideoQA dataset with 21,412 human-generated question-answer samples for comparable transfer of knowledge. Our experiments show that: (i) domain-agnostic knowledge is transferable and (ii) our proposed transfer learning framework can boost VideoQA performance effectively.

Zechen Bai, Yuta Nakashima, Noa Garcia

Have you ever looked at a painting and wondered what is the story behind it? This work presents a framework to bring art closer to people by generating comprehensive descriptions of fine-art paintings. Generating informative descriptions for artworks, however, is extremely challenging, as it requires to 1) describe multiple aspects of the image such as its style, content, or composition, and 2) provide background and contextual knowledge about the artist, their influences, or the historical period. To address these challenges, we introduce a multi-topic and knowledgeable art description framework, which modules the generated sentences according to three artistic topics and, additionally, enhances each description with external knowledge. The framework is validated through an exhaustive analysis, both quantitative and qualitative, as well as a comparative human evaluation, demonstrating outstanding results in terms of both topic diversity and information veracity.

Yiming Qian, Cheikh Brahim El Vaigh, Yuta Nakashima et al.

Buddha statues are a part of human culture, especially of the Asia area, and they have been alongside human civilisation for more than 2,000 years. As history goes by, due to wars, natural disasters, and other reasons, the records that show the built years of Buddha statues went missing, which makes it an immense work for historians to estimate the built years. In this paper, we pursue the idea of building a neural network model that automatically estimates the built years of Buddha statues based only on their face images. Our model uses a loss function that consists of three terms: an MSE loss that provides the basis for built year estimation; a KL divergence-based loss that handles the samples with both an exact built year and a possible range of built years (e.g., dynasty or centuries) estimated by historians; finally a regularisation that utilises both labelled and unlabelled samples based on manifold assumption. By combining those three terms in the training process, we show that our method is able to estimate built years for given images with 37.5 years of a mean absolute error on the test set.

Akihiko Sayo, Diego Thomas, Hiroshi Kawasaki et al.

We propose a new 2D pose refinement network that learns to predict the human bias in the estimated 2D pose. There are biases in 2D pose estimations that are due to differences between annotations of 2D joint locations based on annotators' perception and those defined by motion capture (MoCap) systems. These biases are crafted into publicly available 2D pose datasets and cannot be removed with existing error reduction approaches. Our proposed pose refinement network allows us to efficiently remove the human bias in the estimated 2D poses and achieve highly accurate multi-view 3D human pose estimation.

Yusuke Hirota, Noa Garcia, Mayu Otani et al.

How far can we go with textual representations for understanding pictures? In image understanding, it is essential to use concise but detailed image representations. Deep visual features extracted by vision models, such as Faster R-CNN, are prevailing used in multiple tasks, and especially in visual question answering (VQA). However, conventional deep visual features may struggle to convey all the details in an image as we humans do. Meanwhile, with recent language models' progress, descriptive text may be an alternative to this problem. This paper delves into the effectiveness of textual representations for image understanding in the specific context of VQA. We propose to take description-question pairs as input, instead of deep visual features, and fed them into a language-only Transformer model, simplifying the process and the computational cost. We also experiment with data augmentation techniques to increase the diversity in the training set and avoid learning statistical bias. Extensive evaluations have shown that textual representations require only about a hundred words to compete with deep visual features on both VQA 2.0 and VQA-CP v2.

Cheikh Brahim El Vaigh, Noa Garcia, Benjamin Renoust et al.

The rise of digitization of cultural documents offers large-scale contents, opening the road for development of AI systems in order to preserve, search, and deliver cultural heritage. To organize such cultural content also means to classify them, a task that is very familiar to modern computer science. Contextual information is often the key to structure such real world data, and we propose to use it in form of a knowledge graph. Such a knowledge graph, combined with content analysis, enhances the notion of proximity between artworks so it improves the performances in classification tasks. In this paper, we propose a novel use of a knowledge graph, that is constructed on annotated data and pseudo-labeled data. With label propagation, we boost artwork classification by training a model using a graph convolutional network, relying on the relationships between entities of the knowledge graph. Following a transductive learning framework, our experiments show that relying on a knowledge graph modeling the relations between labeled data and unlabeled data allows to achieve state-of-the-art results on multiple classification tasks on a dataset of paintings, and on a dataset of Buddha statues. Additionally, we show state-of-the-art results for the difficult case of dealing with unbalanced data, with the limitation of disregarding classes with extremely low degrees in the knowledge graph.

Kiichi Goto, Taikan Suehara, Tamaki Yoshioka et al.

Deep learning is a rapidly-evolving technology with possibility to significantly improve physics reach of collider experiments. In this study we developed a novel algorithm of vertex finding for future lepton colliders such as the International Linear Collider. We deploy two networks; one is simple fully-connected layers to look for vertex seeds from track pairs, and the other is a customized Recurrent Neural Network with an attention mechanism and an encoder-decoder structure to associate tracks to the vertex seeds. The performance of the vertex finder is compared with the standard ILC reconstruction algorithm.

Vinay Damodaran, Sharanya Chakravarthy, Akshay Kumar et al.

Visual Question Answering (VQA) is of tremendous interest to the research community with important applications such as aiding visually impaired users and image-based search. In this work, we explore the use of scene graphs for solving the VQA task. We conduct experiments on the GQA dataset which presents a challenging set of questions requiring counting, compositionality and advanced reasoning capability, and provides scene graphs for a large number of images. We adopt image + question architectures for use with scene graphs, evaluate various scene graph generation techniques for unseen images, propose a training curriculum to leverage human-annotated and auto-generated scene graphs, and build late fusion architectures to learn from multiple image representations. We present a multi-faceted study into the use of scene graphs for VQA, making this work the first of its kind.

Bowen Wang, Liangzhi Li, Manisha Verma et al.

Few-shot learning (FSL) approaches are usually based on an assumption that the pre-trained knowledge can be obtained from base (seen) categories and can be well transferred to novel (unseen) categories. However, there is no guarantee, especially for the latter part. This issue leads to the unknown nature of the inference process in most FSL methods, which hampers its application in some risk-sensitive areas. In this paper, we reveal a new way to perform FSL for image classification, using visual representations from the backbone model and weights generated by a newly-emerged explainable classifier. The weighted representations only include a minimum number of distinguishable features and the visualized weights can serve as an informative hint for the FSL process. Finally, a discriminator will compare the representations of each pair of the images in the support set and the query set. Pairs with the highest scores will decide the classification results. Experimental results prove that the proposed method can achieve both good accuracy and satisfactory explainability on three mainstream datasets.

Bowen Wang, Liangzhi Li, Yuta Nakashima et al.

Semantic video segmentation is a key challenge for various applications. This paper presents a new model named Noisy-LSTM, which is trainable in an end-to-end manner, with convolutional LSTMs (ConvLSTMs) to leverage the temporal coherency in video frames. We also present a simple yet effective training strategy, which replaces a frame in video sequence with noises. This strategy spoils the temporal coherency in video frames during training and thus makes the temporal links in ConvLSTMs unreliable, which may consequently improve feature extraction from video frames, as well as serve as a regularizer to avoid overfitting, without requiring extra data annotation or computational costs. Experimental results demonstrate that the proposed model can achieve state-of-the-art performances in both the CityScapes and EndoVis2018 datasets.