Toru Tamaki

Jun Kimata, Tomoya Nitta, Toru Tamaki

In this paper, we propose a data augmentation method for action recognition using instance segmentation. Although many data augmentation methods have been proposed for image recognition, few of them are tailored for action recognition. Our proposed method, ObjectMix, extracts each object region from two videos using instance segmentation and combines them to create new videos. Experiments on two action recognition datasets, UCF101 and HMDB51, demonstrate the effectiveness of the proposed method and show its superiority over VideoMix, a prior work.

Aoi Otani, Ryota Hashiguchi, Kazuki Omi et al.

In the design of action recognition models, the quality of videos is an important issue; however, the trade-off between the quality and performance is often ignored. In general, action recognition models are trained on high-quality videos, hence it is not known how the model performance degrades when tested on low-quality videos, and how much the quality of training videos affects the performance. The issue of video quality is important, however, it has not been studied so far. The goal of this study is to show the trade-off between the performance and the quality of training and test videos by quantitative performance evaluation of several action recognition models for transcoded videos in different qualities. First, we show how the video quality affects the performance of pre-trained models. We transcode the original validation videos of Kinetics400 by changing quality control parameters of JPEG (compression strength) and H.264/AVC (CRF). Then we use the transcoded videos to validate the pre-trained models. Second, we show how the models perform when trained on transcoded videos. We transcode the original training videos of Kinetics400 by changing the quality parameters of JPEG and H.264/AVC. Then we train the models on the transcoded training videos and validate them with the original and transcoded validation videos. Experimental results with JPEG transcoding show that there is no severe performance degradation (up to -1.5%) for compression strength smaller than 70 where no quality degradation is visually observed, and for larger than 80 the performance degrades linearly with respect to the quality index. Experiments with H.264/AVC transcoding show that there is no significant performance loss (up to -1%) with CRF30 while the total size of video files is reduced to 30%.

Tomoya Nitta, Tsubasa Hirakawa, Hironobu Fujiyoshi et al.

In this paper we propose an extension of the Attention Branch Network (ABN) by using instance segmentation for generating sharper attention maps for action recognition. Methods for visual explanation such as Grad-CAM usually generate blurry maps which are not intuitive for humans to understand, particularly in recognizing actions of people in videos. Our proposed method, Object-ABN, tackles this issue by introducing a new mask loss that makes the generated attention maps close to the instance segmentation result. Further the PC loss and multiple attention maps are introduced to enhance the sharpness of the maps and improve the performance of classification. Experimental results with UCF101 and SSv2 shows that the generated maps by the proposed method are much clearer qualitatively and quantitatively than those of the original ABN.

Ryota Hashiguchi, Toru Tamaki

Feature shifts have been shown to be useful for action recognition with CNN-based models since Temporal Shift Module (TSM) was proposed. It is based on frame-wise feature extraction with late fusion, and layer features are shifted along the time direction for the temporal interaction. TokenShift, a recent model based on Vision Transformer (ViT), also uses the temporal feature shift mechanism, which, however, does not fully exploit the structure of Multi-head Self-Attention (MSA) in ViT. In this paper, we propose Multi-head Self/Cross-Attention (MSCA), which fully utilizes the attention structure. TokenShift is based on a frame-wise ViT with features temporally shifted with successive frames (at time t+1 and t-1). In contrast, the proposed MSCA replaces MSA in the frame-wise ViT, and some MSA heads attend to successive frames instead of the current frame. The computation cost is the same as the frame-wise ViT and TokenShift as it simply changes the target to which the attention is taken. There is a choice about which of key, query, and value are taken from the successive frames, then we experimentally compared these variants with Kinetics400. We also investigate other variants in which the proposed MSCA is used along the patch dimension of ViT, instead of the head dimension. Experimental results show that a variant, MSCA-KV, shows the best performance and is better than TokenShift by 0.1% and then ViT by 1.2%.

Shuki Shimizu, Toru Tamaki

In this study, we propose a method for jointly learning of images and videos using a single model. In general, images and videos are often trained by separate models. We propose in this paper a method that takes a batch of images as input to Vision Transformer IV-ViT, and also a set of video frames with temporal aggregation by late fusion. Experimental results on two image datasets and two action recognition datasets are presented.

Taiki Sugiura, Toru Tamaki

Action recognition is a well-established area of research in computer vision. In this paper, we propose S3Aug, a video data augmenatation for action recognition. Unlike conventional video data augmentation methods that involve cutting and pasting regions from two videos, the proposed method generates new videos from a single training video through segmentation and label-to-image transformation. Furthermore, the proposed method modifies certain categories of label images by sampling to generate a variety of videos, and shifts intermediate features to enhance the temporal coherency between frames of the generate videos. Experimental results on the UCF101, HMDB51, and Mimetics datasets demonstrate the effectiveness of the proposed method, paricularlly for out-of-context videos of the Mimetics dataset.

Kenji Matsui, Toru Tamaki, Gwladys Auffret et al.

We propose a feature for action recognition called Trajectory-Set (TS), on top of the improved Dense Trajectory (iDT). The TS feature encodes only trajectories around densely sampled interest points, without any appearance features. Experimental results on the UCF50, UCF101, and HMDB51 action datasets demonstrate that TS is comparable to state-of-the-arts, and outperforms many other methods; for HMDB the accuracy of 85.4%, compared to the best accuracy of 80.2% obtained by a deep method. Our code is available on-line at https://github.com/Gauffret/TrajectorySet .

Shimon Hori, Kazuki Omi, Toru Tamaki

In this paper, we propose a method that extends the query-based object detection model, DETR, to spatio-temporal action detection, which requires maintaining temporal consistency in videos. Our proposed method applies DETR to each frame and uses feature shift to incorporate temporal information. However, DETR's object queries in each frame may correspond to different objects, making a simple feature shift ineffective. To overcome this issue, we propose query matching across different frames, ensuring that queries for the same object are matched and used for the feature shift. Experimental results show that performance on the JHMDB21 dataset improves significantly when query features are shifted using the proposed query matching.

Tomoya Nitta, Takumi Fukuzawa, Toru Tamaki

This paper proposes a method for video captioning that controls the length of generated captions. Previous work on length control often had few levels for expressing length. In this study, we propose two methods of length embedding for fine-grained length control. A traditional embedding method is linear, using a one-hot vector and an embedding matrix. In this study, we propose methods that represent length in multi-hot vectors. One is bit embedding that expresses length in bit representation, and the other is ordinal embedding that uses the binary representation often used in ordinal regression. These length representations of multi-hot vectors are converted into length embedding by a nonlinear MLP. This method allows for not only the length control of caption sentences but also the control of the time when reading the caption. Experiments using ActivityNet Captions and Spoken Moments in Time show that the proposed method effectively controls the length of the generated captions. Analysis of the embedding vectors with ICA shows that length and semantics were learned separately, demonstrating the effectiveness of the proposed embedding methods.

Takumi Fukuzawa, Kensho Hara, Hirokatsu Kataoka et al.

In this paper, we address the issue of static bias in zero-shot action recognition. Action recognition models need to represent the action itself, not the appearance. However, some fully-supervised works show that models often rely on static appearances, such as the background and objects, rather than human actions. This issue, known as static bias, has not been investigated for zero-shot. Although CLIP-based zero-shot models are now common, it remains unclear if they sufficiently focus on human actions, as CLIP primarily captures appearance features related to languages. In this paper, we investigate the influence of static bias in zero-shot action recognition with CLIP-based models. Our approach involves masking backgrounds, objects, and people differently during training and validation. Experiments with masking background show that models depend on background bias as their performance decreases for Kinetics400. However, for Mimetics, which has a weak background bias, masking the background leads to improved performance even if the background is masked during validation. Furthermore, masking both the background and objects in different colors improves performance for SSv2, which has a strong object bias. These results suggest that masking the background or objects during training prevents models from overly depending on static bias and makes them focus more on human action.

Masato Kobayashi, Ning Ding, Toru Tamaki

Action recognition models rely excessively on static cues rather than dynamic human motion, which is known as static bias. This bias leads to poor performance in real-world applications and zero-shot action recognition. In this paper, we propose a method to reduce static bias by separating temporal dynamic information from static scene information. Our approach uses a statistical independence loss between biased and unbiased streams, combined with a scene prediction loss. Our experiments demonstrate that this method effectively reduces static bias and confirm the importance of scene prediction loss.

Ning Ding, Keisuke Fujii, Toru Tamaki

Tactical understanding in badminton involves interpreting not only individual actions but also how tactics are dynamically executed over time. In this paper, we propose \textbf{Shot2Tactic-Caption}, a novel framework for semantic and temporal multi-scale video captioning in badminton, capable of generating shot-level captions that describe individual actions and tactic-level captions that capture how these actions unfold over time within a tactical execution. We also introduce the Shot2Tactic-Caption Dataset, the first badminton captioning dataset containing 5,494 shot captions and 544 tactic captions. Shot2Tactic-Caption adopts a dual-branch design, with both branches including a visual encoder, a spatio-temporal Transformer encoder, and a Transformer-based decoder to generate shot and tactic captions. To support tactic captioning, we additionally introduce a Tactic Unit Detector that identifies valid tactic units, tactic types, and tactic states (e.g., Interrupt, Resume). For tactic captioning, we further incorporate a shot-wise prompt-guided mechanism, where the predicted tactic type and state are embedded as prompts and injected into the decoder via cross-attention. The shot-wise prompt-guided mechanism enables our system not only to describe successfully executed tactics but also to capture tactical executions that are temporarily interrupted and later resumed. Experimental results demonstrate the effectiveness of our framework in generating both shot and tactic captions. Ablation studies show that the ResNet50-based spatio-temporal encoder outperforms other variants, and that shot-wise prompt structuring leads to more coherent and accurate tactic captioning.

Qiuyu Chen, Xin Jin, Yue Song et al.

This paper reviews the 1st International Workshop on Disentangled Representation Learning for Controllable Generation (DRL4Real), held in conjunction with ICCV 2025. The workshop aimed to bridge the gap between the theoretical promise of Disentangled Representation Learning (DRL) and its application in realistic scenarios, moving beyond synthetic benchmarks. DRL4Real focused on evaluating DRL methods in practical applications such as controllable generation, exploring advancements in model robustness, interpretability, and generalization. The workshop accepted 9 papers covering a broad range of topics, including the integration of novel inductive biases (e.g., language), the application of diffusion models to DRL, 3D-aware disentanglement, and the expansion of DRL into specialized domains like autonomous driving and EEG analysis. This summary details the workshop's objectives, the themes of the accepted papers, and provides an overview of the methodologies proposed by the authors.

Kazuki Omi, Jion Oshima, Toru Tamaki

This paper proposes a method for spatio-temporal action detection (STAD) that directly generates action tubes from the original video without relying on post-processing steps such as IoU-based linking and clip splitting. Our approach applies query-based detection (DETR) to each frame and matches DETR queries to link the same person across frames. We introduce the Query Matching Module (QMM), which uses metric learning to bring queries for the same person closer together across frames compared to queries for different people. Action classes are predicted using the sequence of queries obtained from QMM matching, allowing for variable-length inputs from videos longer than a single clip. Experimental results on JHMDB, UCF101-24, and AVA datasets demonstrate that our method performs well for large position changes of people while offering superior computational efficiency and lower resource requirements.

Kodai Kamiya, Toru Tamaki

We propose a new method for learning videos by aggregating multiple models by sequentially extracting video clips from untrimmed video. The proposed method reduces the correlation between clips by feeding clips to multiple models in turn and synchronizes these models through federated learning. Experimental results show that the proposed method improves the performance compared to the no synchronization.

Kento Terao, Toru Tamaki, Bisser Raytchev et al.

We propose a novel approach to identify the difficulty of visual questions for Visual Question Answering (VQA) without direct supervision or annotations to the difficulty. Prior works have considered the diversity of ground-truth answers of human annotators. In contrast, we analyze the difficulty of visual questions based on the behavior of multiple different VQA models. We propose to cluster the entropy values of the predicted answer distributions obtained by three different models: a baseline method that takes as input images and questions, and two variants that take as input images only and questions only. We use a simple k-means to cluster the visual questions of the VQA v2 validation set. Then we use state-of-the-art methods to determine the accuracy and the entropy of the answer distributions for each cluster. A benefit of the proposed method is that no annotation of the difficulty is required, because the accuracy of each cluster reflects the difficulty of visual questions that belong to it. Our approach can identify clusters of difficult visual questions that are not answered correctly by state-of-the-art methods. Detailed analysis on the VQA v2 dataset reveals that 1) all methods show poor performances on the most difficult cluster (about 10\% accuracy), 2) as the cluster difficulty increases, the answers predicted by the different methods begin to differ, and 3) the values of cluster entropy are highly correlated with the cluster accuracy. We show that our approach has the advantage of being able to assess the difficulty of visual questions without ground-truth (\ie, the test set of VQA v2) by assigning them to one of the clusters. We expect that this can stimulate the development of novel directions of research and new algorithms.

Kento Terao, Toru Tamaki, Bisser Raytchev et al.

Visual question answering (VQA) is a task of answering a visual question that is a pair of question and image. Some visual questions are ambiguous and some are clear, and it may be appropriate to change the ambiguity of questions from situation to situation. However, this issue has not been addressed by any prior work. We propose a novel task, rephrasing the questions by controlling the ambiguity of the questions. The ambiguity of a visual question is defined by the use of the entropy of the answer distribution predicted by a VQA model. The proposed model rephrases a source question given with an image so that the rephrased question has the ambiguity (or entropy) specified by users. We propose two learning strategies to train the proposed model with the VQA v2 dataset, which has no ambiguity information. We demonstrate the advantage of our approach that can control the ambiguity of the rephrased questions, and an interesting observation that it is harder to increase than to reduce ambiguity.

Zhao Fangda, Toru Tamaki, Takio Kurita et al.

We propose an easy-to-use non-overlapping camera calibration method. First, successive images are fed to a PoseNet-based network to obtain ego-motion of cameras between frames. Next, the pose between cameras are estimated. Instead of using a batch method, we propose an on-line method of the inter-camera pose estimation. Furthermore, we implement the entire procedure on a computation graph. Experiments with simulations and the KITTI dataset show the proposed method to be effective in simulation.

Daisuke Ogawa, Toru Tamaki, Tsubasa Hirakawa et al.

An efficient inverse reinforcement learning for generating trajectories is proposed based of 2D and 3D activity forecasting. We modify reward function with $L_p$ norm and propose convolution into value iteration steps, which is called convolutional value iteration. Experimental results with seabird trajectories (43 for training and 10 for test), our method is best in terms of MHD error and performs fastest. Generated trajectories for interpolating missing parts of trajectories look much similar to real seabird trajectories than those by the previous works.

Shohei Hayashi, Bisser Raytchev, Toru Tamaki et al.

In this paper we propose a novel deep learning-based algorithm for biomedical image segmentation which uses a sequential attention mechanism able to shift the focus of attention across the image in a selective way, allowing subareas which are more difficult to classify to be processed at increased resolution. The spatial distribution of class information in each subarea is learned using a retina-like representation where resolution decreases with distance from the center of attention. The final segmentation is achieved by averaging class predictions over overlapping subareas, utilizing the power of ensemble learning to increase segmentation accuracy. Experimental results for semantic segmentation task for which only a few training images are available show that a CNN using the proposed method outperforms both a patch-based classification CNN and a fully convolutional-based method.

Tsubasa Hirakawa, Takayoshi Yamashita, Toru Tamaki et al.

Path prediction is a fundamental task for estimating how pedestrians or vehicles are going to move in a scene. Because path prediction as a task of computer vision uses video as input, various information used for prediction, such as the environment surrounding the target and the internal state of the target, need to be estimated from the video in addition to predicting paths. Many prediction approaches that include understanding the environment and the internal state have been proposed. In this survey, we systematically summarize methods of path prediction that take video as input and and extract features from the video. Moreover, we introduce datasets used to evaluate path prediction methods quantitatively.

Daisuke Ogawa, Toru Tamaki, Bisser Raytchev et al.

In many cases, such as trajectories clustering and classification, we often divide a trajectory into segments as preprocessing. In this paper, we propose a trajectory semantic segmentation method based on learned behavior models. In the proposed method, we learn some behavior models from video sequences. Next, using learned behavior models and a hidden Markov model, we segment a trajectory into semantic segments. Comparing with the Ramer-Douglas-Peucker algorithm, we show the effectiveness of the proposed method.

Tsubasa Hirakawa, Toru Tamaki, Bisser Raytchev et al.

Colorectal endoscopy is important for the early detection and treatment of colorectal cancer and is used worldwide. A computer-aided diagnosis (CAD) system that provides an objective measure to endoscopists during colorectal endoscopic examinations would be of great value. In this study, we describe a newly developed CAD system that provides real-time objective measures. Our system captures the video stream from an endoscopic system and transfers it to a desktop computer. The captured video stream is then classified by a pretrained classifier and the results are displayed on a monitor. The experimental results show that our developed system works efficiently in actual endoscopic examinations and is medically significant.

Toru Tamaki, Shoji Sonoyama, Takio Kurita et al.

This paper proposes a method for domain adaptation that extends the maximum margin domain transfer (MMDT) proposed by Hoffman et al., by introducing L2 distance constraints between samples of different domains; thus, our method is denoted as MMDTL2. Motivated by the differences between the images taken by narrow band imaging (NBI) endoscopic devices, we utilize different NBI devices as different domains and estimate the transformations between samples of different domains, i.e., image samples taken by different NBI endoscope systems. We first formulate the problem in the primal form, and then derive the dual form with much lesser computational costs as compared to the naive approach. From our experimental results using NBI image datasets from two different NBI endoscopic devices, we find that MMDTL2 is better than MMDT and also support vector machines without adaptation, especially when NBI image features are high-dimensional and the per-class training samples are greater than 20.

Shoji Sonoyama, Toru Tamaki, Tsubasa Hirakawa et al.

In this paper we propose a method for transfer learning of endoscopic images. For transferring between features obtained from images taken by different (old and new) endoscopes, we extend the Max-Margin Domain Transfer (MMDT) proposed by Hoffman et al. in order to use L2 distance constraints as regularization, called Max-Margin Domain Transfer with L2 Distance Constraints (MMDTL2). Furthermore, we develop the dual formulation of the optimization problem in order to reduce the computation cost. Experimental results demonstrate that the proposed MMDTL2 outperforms MMDT for real data sets taken by different endoscopes.

Toru Tamaki, Shoji Sonoyama, Tsubasa Hirakawa et al.

In this paper we report results for recognizing colorectal NBI endoscopic images by using features extracted from convolutional neural network (CNN). In this comparative study, we extract features from different layers from different CNN models, and then train linear SVM classifiers. Experimental results with 10-fold cross validations show that features from first few convolution layers are enough to achieve similar performance (i.e., recognition rate of 95%) with non-CNN local features such as Bag-of-Visual words, Fisher vector, and VLAD.