Yusuke Matsui

Yingxuan Li, Kiyoharu Aizawa, Yusuke Matsui

The expanding market for e-comics has spurred interest in the development of automated methods to analyze comics. For further understanding of comics, an automated approach is needed to link text in comics to characters speaking the words. Comics speaker detection research has practical applications, such as automatic character assignment for audiobooks, automatic translation according to characters' personalities, and inference of character relationships and stories. To deal with the problem of insufficient speaker-to-text annotations, we created a new annotation dataset Manga109Dialog based on Manga109. Manga109Dialog is the world's largest comics speaker annotation dataset, containing 132,692 speaker-to-text pairs. We further divided our dataset into different levels by prediction difficulties to evaluate speaker detection methods more appropriately. Unlike existing methods mainly based on distances, we propose a deep learning-based method using scene graph generation models. Due to the unique features of comics, we enhance the performance of our proposed model by considering the frame reading order. We conducted experiments using Manga109Dialog and other datasets. Experimental results demonstrate that our scene-graph-based approach outperforms existing methods, achieving a prediction accuracy of over 75%.

Kunato Nishina, Yusuke Matsui

Text-to-image models have shown progress in recent years. Along with this progress, generating vector graphics from text has also advanced. SVG is a popular format for vector graphics, and SVG represents a scene with XML text. Therefore, Large Language Models can directly process SVG code. Taking this into account, we focused on editing SVG with LLMs. For quantitative evaluation of LLMs' ability to edit SVG, we propose SVGEditBench. SVGEditBench is a benchmark for assessing the LLMs' ability to edit SVG code. We also show the GPT-4 and GPT-3.5 results when evaluated on the proposed benchmark. In the experiments, GPT-4 showed superior performance to GPT-3.5 both quantitatively and qualitatively. The dataset is available at https://github.com/mti-lab/SVGEditBench.

Yusuke Matsui

Approximate nearest neighbor search (ANNS) is an essential building block for applications like RAG but can sometimes yield results that are overly similar to each other. In certain scenarios, search results should be similar to the query and yet diverse. We propose LotusFilter, a post-processing module to diversify ANNS results. We precompute a cutoff table summarizing vectors that are close to each other. During the filtering, LotusFilter greedily looks up the table to delete redundant vectors from the candidates. We demonstrated that the LotusFilter operates fast (0.02 [ms/query]) in settings resembling real-world RAG applications, utilizing features such as OpenAI embeddings. Our code is publicly available at https://github.com/matsui528/lotf.

Yusuke Matsui, Tatsuya Yokota

We propose a new operator defined between two tensors, the broadcast product. The broadcast product calculates the Hadamard product after duplicating elements to align the shapes of the two tensors. Complex tensor operations in libraries like \texttt{numpy} can be succinctly represented as mathematical expressions using the broadcast product. Finally, we propose a novel tensor decomposition using the broadcast product, highlighting its potential applications in dimensionality reduction.

Jeonghun Baek, Yusuke Matsui, Kiyoharu Aizawa

Scene text recognition (STR) task has a common practice: All state-of-the-art STR models are trained on large synthetic data. In contrast to this practice, training STR models only on fewer real labels (STR with fewer labels) is important when we have to train STR models without synthetic data: for handwritten or artistic texts that are difficult to generate synthetically and for languages other than English for which we do not always have synthetic data. However, there has been implicit common knowledge that training STR models on real data is nearly impossible because real data is insufficient. We consider that this common knowledge has obstructed the study of STR with fewer labels. In this work, we would like to reactivate STR with fewer labels by disproving the common knowledge. We consolidate recently accumulated public real data and show that we can train STR models satisfactorily only with real labeled data. Subsequently, we find simple data augmentation to fully exploit real data. Furthermore, we improve the models by collecting unlabeled data and introducing semi- and self-supervised methods. As a result, we obtain a competitive model to state-of-the-art methods. To the best of our knowledge, this is the first study that 1) shows sufficient performance by only using real labels and 2) introduces semi- and self-supervised methods into STR with fewer labels. Our code and data are available: https://github.com/ku21fan/STR-Fewer-Labels

Takuto Onikubo, Yusuke Matsui

Recently, text-to-image generative models have been misused to create unauthorized malicious images of individuals, posing a growing social problem. Previous solutions, such as Anti-DreamBooth, add adversarial noise to images to protect them from being used as training data for malicious generation. However, we found that the adversarial noise can be removed by adversarial purification methods such as DiffPure. Therefore, we propose a new adversarial attack method that adds strong perturbation on the high-frequency areas of images to make it more robust to adversarial purification. Our experiment showed that the adversarial images retained noise even after adversarial purification, hindering malicious image generation.

Yingxuan Li, Ryota Hinami, Kiyoharu Aizawa et al.

Recognizing characters and predicting speakers of dialogue are critical for comic processing tasks, such as voice generation or translation. However, because characters vary by comic title, supervised learning approaches like training character classifiers which require specific annotations for each comic title are infeasible. This motivates us to propose a novel zero-shot approach, allowing machines to identify characters and predict speaker names based solely on unannotated comic images. In spite of their importance in real-world applications, these task have largely remained unexplored due to challenges in story comprehension and multimodal integration. Recent large language models (LLMs) have shown great capability for text understanding and reasoning, while their application to multimodal content analysis is still an open problem. To address this problem, we propose an iterative multimodal framework, the first to employ multimodal information for both character identification and speaker prediction tasks. Our experiments demonstrate the effectiveness of the proposed framework, establishing a robust baseline for these tasks. Furthermore, since our method requires no training data or annotations, it can be used as-is on any comic series.

Ryoya Nara, Yu-Chieh Lin, Yuji Nozawa et al.

Many image retrieval studies use metric learning to train an image encoder. However, metric learning cannot handle differences in users' preferences, and requires data to train an image encoder. To overcome these limitations, we revisit relevance feedback, a classic technique for interactive retrieval systems, and propose an interactive CLIP-based image retrieval system with relevance feedback. Our retrieval system first executes the retrieval, collects each user's unique preferences through binary feedback, and returns images the user prefers. Even when users have various preferences, our retrieval system learns each user's preference through the feedback and adapts to the preference. Moreover, our retrieval system leverages CLIP's zero-shot transferability and achieves high accuracy without training. We empirically show that our retrieval system competes well with state-of-the-art metric learning in category-based image retrieval, despite not training image encoders specifically for each dataset. Furthermore, we set up two additional experimental settings where users have various preferences: one-label-based image retrieval and conditioned image retrieval. In both cases, our retrieval system effectively adapts to each user's preferences, resulting in improved accuracy compared to image retrieval without feedback. Overall, our work highlights the potential benefits of integrating CLIP with classic relevance feedback techniques to enhance image retrieval.

Yutaro Oguri, Yusuke Matsui

We present a theoretical and empirical analysis of the adaptive entry point selection for graph-based approximate nearest neighbor search (ANNS). We introduce novel concepts: $b\textit{-monotonic path}$ and $B\textit{-MSNET}$, which better capture an actual graph in practical algorithms than existing concepts like MSNET. We prove that adaptive entry point selection offers better performance upper bound than the fixed central entry point under more general conditions than previous work. Empirically, we validate the method's effectiveness in accuracy, speed, and memory usage across various datasets, especially in challenging scenarios with out-of-distribution data and hard instances. Our comprehensive study provides deeper insights into optimizing entry points for graph-based ANNS for real-world high-dimensional data applications.

Yutaro Oguri, Mai Nishimura, Yusuke Matsui

We present the first systematic investigation of graph reordering effects for graph-based Approximate Nearest Neighbor Search (ANNS) on a GPU. While graph-based ANNS has become the dominant paradigm for modern AI applications, recent approaches focus on algorithmic innovations while neglecting memory layout considerations that significantly affect execution time. Our unified evaluation framework enables comprehensive evaluation of diverse reordering strategies across different graph indices through a graph adapter that converts arbitrary graph topologies into a common representation and a GPU-optimized graph traversal engine. We conduct a comprehensive analysis across diverse datasets and state-of-the-art graph indices, introducing analysis metrics that quantify the relationship between structural properties and memory layout effectiveness. Our GPU-targeted reordering achieves up to 15$\%$ QPS improvements while preserving search accuracy, demonstrating that memory layout optimization operates orthogonally to existing algorithmic innovations. We will release all code upon publication to facilitate reproducibility and foster further research.

Atsuki Sato, Yusuke Matsui

Recent studies have demonstrated that learned Bloom filters, which combine machine learning with the classical Bloom filter, can achieve superior memory efficiency. However, existing learned Bloom filters face two critical unresolved challenges: the balance between the machine learning model size and the Bloom filter size is not optimal, and the reject time cannot be minimized effectively. We propose the Cascaded Learned Bloom Filter (CLBF) to address these issues. Our dynamic programming-based optimization automatically selects configurations that achieve an optimal balance between the model and filter sizes while minimizing reject time. Experiments on real-world datasets show that CLBF reduces memory usage by up to 24% and decreases reject time by up to 14 times compared to state-of-the-art learned Bloom filters.

Yutaro Oguri, Yusuke Matsui

Despite the efficacy of graph-based algorithms for Approximate Nearest Neighbor (ANN) searches, the optimal tuning of such systems remains unclear. This study introduces a method to tune the performance of off-the-shelf graph-based indexes, focusing on the dimension of vectors, database size, and entry points of graph traversal. We utilize a black-box optimization algorithm to perform integrated tuning to meet the required levels of recall and Queries Per Second (QPS). We applied our approach to Task A of the SISAP 2023 Indexing Challenge and got second place in the 10M and 30M tracks. It improves performance substantially compared to brute force methods. This research offers a universally applicable tuning method for graph-based indexes, extending beyond the specific conditions of the competition to broader uses.

Yoichiro Hisadome, Yusuke Matsui

We propose a method for speeding up a 3D point cloud registration through a cascading feature extraction. The current approach with the highest accuracy is realized by iteratively executing feature extraction and registration using deep features. However, iterative feature extraction takes time. Our proposed method significantly reduces the computational cost using cascading shallow layers. Our idea is to omit redundant computations that do not always contribute to the final accuracy. The proposed approach is approximately three times faster than the existing methods without a loss of accuracy.

Ryota Hinami, Shonosuke Ishiwatari, Kazuhiko Yasuda et al.

We tackle the problem of machine translation of manga, Japanese comics. Manga translation involves two important problems in machine translation: context-aware and multimodal translation. Since text and images are mixed up in an unstructured fashion in Manga, obtaining context from the image is essential for manga translation. However, it is still an open problem how to extract context from image and integrate into MT models. In addition, corpus and benchmarks to train and evaluate such model is currently unavailable. In this paper, we make the following four contributions that establishes the foundation of manga translation research. First, we propose multimodal context-aware translation framework. We are the first to incorporate context information obtained from manga image. It enables us to translate texts in speech bubbles that cannot be translated without using context information (e.g., texts in other speech bubbles, gender of speakers, etc.). Second, for training the model, we propose the approach to automatic corpus construction from pairs of original manga and their translations, by which large parallel corpus can be constructed without any manual labeling. Third, we created a new benchmark to evaluate manga translation. Finally, on top of our proposed methods, we devised a first comprehensive system for fully automated manga translation.

Fan Yang, Ryota Hinami, Yusuke Matsui et al.

Diffusion is commonly used as a ranking or re-ranking method in retrieval tasks to achieve higher retrieval performance, and has attracted lots of attention in recent years. A downside to diffusion is that it performs slowly in comparison to the naive k-NN search, which causes a non-trivial online computational cost on large datasets. To overcome this weakness, we propose a novel diffusion technique in this paper. In our work, instead of applying diffusion to the query, we pre-compute the diffusion results of each element in the database, making the online search a simple linear combination on top of the k-NN search process. Our proposed method becomes 10~ times faster in terms of online search speed. Moreover, we propose to use late truncation instead of early truncation in previous works to achieve better retrieval performance.

Yusuke Matsui, Ryota Hinami, Shin'ichi Satoh

Existing approximate nearest neighbor search systems suffer from two fundamental problems that are of practical importance but have not received sufficient attention from the research community. First, although existing systems perform well for the whole database, it is difficult to run a search over a subset of the database. Second, there has been no discussion concerning the performance decrement after many items have been newly added to a system. We develop a reconfigurable inverted index (Rii) to resolve these two issues. Based on the standard IVFADC system, we design a data layout such that items are stored linearly. This enables us to efficiently run a subset search by switching the search method to a linear PQ scan if the size of a subset is small. Owing to the linear layout, the data structure can be dynamically adjusted after new items are added, maintaining the fast speed of the system. Extensive comparisons show that Rii achieves a comparable performance with state-of-the art systems such as Faiss.

Toru Ogawa, Atsushi Otsubo, Rei Narita et al.

With the growth of digitized comics, image understanding techniques are becoming important. In this paper, we focus on object detection, which is a fundamental task of image understanding. Although convolutional neural networks (CNN)-based methods archived good performance in object detection for naturalistic images, there are two problems in applying these methods to the comic object detection task. First, there is no large-scale annotated comics dataset. The CNN-based methods require large-scale annotations for training. Secondly, the objects in comics are highly overlapped compared to naturalistic images. This overlap causes the assignment problem in the existing CNN-based methods. To solve these problems, we proposed a new annotation dataset and a new CNN model. We annotated an existing image dataset of comics and created the largest annotation dataset, named Manga109-annotations. For the assignment problem, we proposed a new CNN-based detector, SSD300-fork. We compared SSD300-fork with other detection methods using Manga109-annotations and confirmed that our model outperformed them based on the mAP score.

Yasunori Kudo, Keisuke Ogaki, Yusuke Matsui et al.

The task of three-dimensional (3D) human pose estimation from a single image can be divided into two parts: (1) Two-dimensional (2D) human joint detection from the image and (2) estimating a 3D pose from the 2D joints. Herein, we focus on the second part, i.e., a 3D pose estimation from 2D joint locations. The problem with existing methods is that they require either (1) a 3D pose dataset or (2) 2D joint locations in consecutive frames taken from a video sequence. We aim to solve these problems. For the first time, we propose a method that learns a 3D human pose without any 3D datasets. Our method can predict a 3D pose from 2D joint locations in a single image. Our system is based on the generative adversarial networks, and the networks are trained in an unsupervised manner. Our primary idea is that, if the network can predict a 3D human pose correctly, the 3D pose that is projected onto a 2D plane should not collapse even if it is rotated perpendicularly. We evaluated the performance of our method using Human3.6M and the MPII dataset and showed that our network can predict a 3D pose well even if the 3D dataset is not available during training.

Ryota Hinami, Yusuke Matsui, Shin'ichi Satoh

Region-based image retrieval (RBIR) technique is revisited. In early attempts at RBIR in the late 90s, researchers found many ways to specify region-based queries and spatial relationships; however, the way to characterize the regions, such as by using color histograms, were very poor at that time. Here, we revisit RBIR by incorporating semantic specification of objects and intuitive specification of spatial relationships. Our contributions are the following. First, to support multiple aspects of semantic object specification (category, instance, and attribute), we propose a multitask CNN feature that allows us to use deep learning technique and to jointly handle multi-aspect object specification. Second, to help users specify spatial relationships among objects in an intuitive way, we propose recommendation techniques of spatial relationships. In particular, by mining the search results, a system can recommend feasible spatial relationships among the objects. The system also can recommend likely spatial relationships by assigned object category names based on language prior. Moreover, object-level inverted indexing supports very fast shortlist generation, and re-ranking based on spatial constraints provides users with instant RBIR experiences.

Yusuke Matsui, Keisuke Ogaki, Toshihiko Yamasaki et al.

Data clustering is a fundamental operation in data analysis. For handling large-scale data, the standard k-means clustering method is not only slow, but also memory-inefficient. We propose an efficient clustering method for billion-scale feature vectors, called PQk-means. By first compressing input vectors into short product-quantized (PQ) codes, PQk-means achieves fast and memory-efficient clustering, even for high-dimensional vectors. Similar to k-means, PQk-means repeats the assignment and update steps, both of which can be performed in the PQ-code domain. Experimental results show that even short-length (32 bit) PQ-codes can produce competitive results compared with k-means. This result is of practical importance for clustering in memory-restricted environments. Using the proposed PQk-means scheme, the clustering of one billion 128D SIFT features with K = 10^5 is achieved within 14 hours, using just 32 GB of memory consumption on a single computer.

Yusuke Matsui, Toshihiko Yamasaki, Kiyoharu Aizawa

In this paper, we propose a product quantization table (PQTable); a fast search method for product-quantized codes via hash-tables. An identifier of each database vector is associated with the slot of a hash table by using its PQ-code as a key. For querying, an input vector is PQ-encoded and hashed, and the items associated with that code are then retrieved. The proposed PQTable produces the same results as a linear PQ scan, and is 10^2 to 10^5 times faster. Although state-of-the-art performance can be achieved by previous inverted-indexing-based approaches, such methods require manually-designed parameter setting and significant training; our PQTable is free of these limitations, and therefore offers a practical and effective solution for real-world problems. Specifically, when the vectors are highly compressed, our PQTable achieves one of the fastest search performances on a single CPU to date with significantly efficient memory usage (0.059 ms per query over 10^9 data points with just 5.5 GB memory consumption). Finally, we show that our proposed PQTable can naturally handle the codes of an optimized product quantization (OPQTable).

Yusuke Matsui, Kota Ito, Yuji Aramaki et al.

Manga (Japanese comics) are popular worldwide. However, current e-manga archives offer very limited search support, including keyword-based search by title or author, or tag-based categorization. To make the manga search experience more intuitive, efficient, and enjoyable, we propose a content-based manga retrieval system. First, we propose a manga-specific image-describing framework. It consists of efficient margin labeling, edge orientation histogram feature description, and approximate nearest-neighbor search using product quantization. Second, we propose a sketch-based interface as a natural way to interact with manga content. The interface provides sketch-based querying, relevance feedback, and query retouch. For evaluation, we built a novel dataset of manga images, Manga109, which consists of 109 comic books of 21,142 pages drawn by professional manga artists. To the best of our knowledge, Manga109 is currently the biggest dataset of manga images available for research. We conducted a comparative study, a localization evaluation, and a large-scale qualitative study. From the experiments, we verified that: (1) the retrieval accuracy of the proposed method is higher than those of previous methods; (2) the proposed method can localize an object instance with reasonable runtime and accuracy; and (3) sketch querying is useful for manga search.