論文

Active learning refers to label-efficient algorithms that use the most representative samples for labeling when creating training data. In this paper, we propose a model that derives the most informative unlabeled samples from the output of a task model. The tasks arc a classification problem, multi-label classification and a semantic segmentation problem. The model consists of an uncertainty indicator generator and a task model. After training the task model with labeled samples, the model predicts unlabeled samples, and based on the prediction results, the uncertainty indicator generator outputs an uncertainty indicator for each unlabeled sample. Samples with a higher uncertainty indicator are considered to be more informative and selected. As a result of experiments using multiple datasets, our model achieved better accuracy than conventional active learning methods and reduced execution time by a factor of 10.

Perceptual loss, calculated by VGG network pre-trained on ImageNet, has been widely employed in the past for super-resolution tasks, enabling the generation of photo-realistic images. However, it has been reported that grid-like artifacts frequently appear in the generated images. To address this problem, we consider that large-scale pre-trained models can make significant contributions to super-resolution across different scenes. In particular, by combining language, those models can exhibit a strong capability to comprehend complex scenes, potentially enhancing super-resolution performance. Therefore, this paper proposes new perceptual loss with Contrastive Language-Image Pre-training (CLIP) based on Vision Transformer (ViT) instead of VGG network. The results demonstrate our proposed perceptual loss can generate photorealistic images without grid-like artifacts.

In outdoor environments, Vision-and-Language Navigation (VLN) requires an agent to rely on multi-modal cues from real-world urban environments and natural language instructions. While existing outdoor VLN models predict actions using a combination of panorama and instruction features, this approach ignores objects in the environment and learns data bias to fail navigation. According to our preliminary findings, most instances of navigation failure in previous models were due to turning or stopping at the wrong place. In contrast, humans intuitively frequently use identifiable objects or store names as reference landmarks, ensuring accurate turns and stops, especially in unfamiliar places. To address this insight gap, we propose an Object-Attention VLN (OAVLN) model that helps the agent focus on relevant objects during training and understand the environment better. Our model outperforms previous methods in all evaluation metrics under both seen and unseen scenarios on two existing benchmark datasets, Touchdown and map2seq.

In this paper, we propose a novel formula-driven supervised learning (FDSL) framework for pre-training an environmental sound analysis model by leveraging acoustic signals parametrically synthesized through formula-driven methods. Specifically, we outline detailed procedures and evaluate their effectiveness for sound event detection (SED). The SED task, which involves estimating the types and timings of sound events, is particularly challenged by the difficulty of acquiring a sufficient quantity of accurately labeled training data. Moreover, it is well known that manually annotated labels often contain noises and are significantly influenced by the subjective judgment of annotators. To address these challenges, we propose a novel pre-training method that utilizes a synthetic dataset, Formula-SED, where acoustic data are generated solely based on mathematical formulas. The proposed method enables large-scale pre-training by using the synthesis parameters applied at each time step as ground truth labels, thereby eliminating label noise and bias. We demonstrate that large-scale pre-training with Formula-SED significantly enhances model accuracy and accelerates training, as evidenced by our results in the DESED dataset used for DCASE2023 Challenge Task 4. The project page is at https://yutoshibata07.github.io/Formula-SED/.

Interaction recognition between multiple individuals is crucial in many practical scenarios, particularly in sports analysis, yet remains challenging due to complex dynamics and directional ambiguity. Most existing methods primarily focus on single-person actions, which are insufficient for analyzing multi-person interactions. The SportsHHI dataset, designed specifically for human-human interactions in sports videos, provides detailed annotations but faces limitations from strict subject-object direction definitions, class imbalance, and inadequate encoding of relative spatial information. To overcome these issues, this study proposes three enhancements: (1) redefining relative spatial encoding to include global positional context, (2) introducing a directionality-agnostic evaluation method suitable for bidirectional interactions, and (3) employing focal loss to address class imbalance. Experimental results on SportsHHI demonstrate the effectiveness of the proposed improvements, achieving up to an increase 2.67 at mAP compared to baseline methods.

Adolescent Idiopathic Scoliosis (AIS) is a common form of scoliosis found among adolescents. The abnormal spinal curvature along with twisting of the vertebrae requires early detection before the vertebrae alignment worsens. Furthermore, surgical treatment is crucial for those with severe conditions. In Japan, annual scoliosis screening is conducted during school health checkups. These screening procedures often involve an initial screening without the use of ionizing radiation (such as X-rays) and an X-ray screening later for those suspected with scoliosis. Currently, X-ray examinations produce the most accurate results for detecting the spinal position. However, noninvasive alternative methods, such as Moire imaging, provide satisfactory initial results. Unfortunately, the production of Moire cameras have halted in recent years. This paper proposes an alternative, non-invasive and non-ionizing radioactive method to detect spinal alignment. Depth images provide an extra dimension of information compared to RGB images. The proposed method utilizes this extra dimension of depth values to create different types of images that are then trained using Convolutional Neural Networks to predict the spinal alignment. Results indicate that Moire images reproduced from depth images produce the best spinal alignment. However, other types of images derived from the depth image also have higher accuracy when compared with the depth image itself.

Osteoarthritis (OA) describes a degenerative joint disorder that is prevalent among older people and typically results in swollen and inflamed joints. The aim of this paper is to develop a method using images, videos and thermal data of 100 patients taken at Keio University Hospital to detect OA in hands. By using hand pose estimation on the video data, joint angles can be calculated and subsequently transformed into feature vectors. For the thermal and RGB images, hand keypoint detectors were trained to identify and crop the appropriate joints within the images. The resulting extracted features are combined and further trained on Support Vector Machines and Convolutional Neural Networks to obtain the final binary classification for each joint. While the proposed method generally shows favorable accuracy and F1-scores on the Proximal (PIP) and Distal Interphalangeal (DIP) joints, the performance on the Metacarpophalangeal (MCP) joints is limited by the low occurrence of affected joints in the dataset. We further compare the different modalities and found that, apart from the combined approach, using video data provides the best results. Clinical Relevance - The proposed method shows promising first results for the usage of visual and thermal data in combination with machine learning in order to detect OA

Shadow removal for document images is a major task for digitized document applications. Recent shadow removal models have been trained on pairs of shadow images and shadow-free images. However, obtaining a large-scale and diverse dataset is laborious and remains a great challenge. Thus, only small real datasets are available. To create relatively large datasets, a graphic renderer has been used to synthesize shadows, nonetheless, it is still necessary to capture real documents. Thus, the number of unique documents is limited, which negatively affects a network's performance. In this paper, we present a large-scale and diverse dataset called fully synthetic document shadow removal dataset (FSDSRD) that does not require capturing documents. The experiments showed that the networks (pre-)trained on FSDSRD provided better results than networks trained only on real datasets. Additionally, because foreground maps are available in our dataset, we leveraged them during training for multitask learning, which provided noticeable improvements. The code is available at: https://github.com/IsHYuhi/DSRFGD.

Pre-training video transformers generally requires a large amount of data, presenting significant challenges in terms of data collection costs and concerns related to privacy, licensing, and inherent biases. Synthesizing data is one of the promising ways to solve these issues, yet pre-training solely on synthetic data has its own challenges. In this paper, we introduce an effective self-supervised learning framework for videos that leverages readily available and less costly static images. Specifically, we define the Pseudo Motion Generator (PMG) module that recursively applies image transformations to generate pseudo-motion videos from images. These pseudo-motion videos are then leveraged in masked video modeling. Our approach is applicable to synthetic images as well, thus entirely freeing video pre-training from data collection costs and other concerns in real data. Through experiments in action recognition tasks, we demonstrate that this framework allows effective learning of spatio-temporal features through pseudo-motion videos, significantly improving over existing methods which also use static images and partially outperforming those using both real and synthetic videos. These results uncover fragments of what video transformers learn through masked video modeling.

In order to train a classifier for self checkout system in convenience store at low cost, a few-shot domain adaptation problem has to be solved. Since the system treats a classifier for large number of classes, conventional method of few-shot domain adaptation should be extended for many classes. This paper propose to exploit meta-class information by executing the adaptation on the normal-class level and the meta-class level simultaneously. The proposed method are shown to be effective for improving adaptation accuracy of a classifier for many classes. The results of our ablation study implies that i) the meta-class should be decided by using k-means clustering method rather than clustering manually, and that ii) the ratio between the number of normal-class and the number of meta-class should be fixed.

Real-time 3D trajectory player tracking in sports plays a crucial role in tactical analysis, performance evaluation, and enhancing spectator experience. Traditional systems rely on multi-camera setups, but are constrained by the inherently two-dimensional nature of video data and the need for complex 3D reconstruction processing, making real-time analysis challenging. Basketball, in particular, represents one of the most difficult scenarios in the MOT field, as ten players move rapidly and complexly within a confined court space, with frequent occlusions caused by intense physical contact. To address these challenges, this paper constructs BasketLiDAR, the first multimodal dataset in the sports MOT field that combines LiDAR point clouds with synchronized multi-view camera footage in a professional basketball environment, and proposes a novel MOT framework that simultaneously achieves improved tracking accuracy and reduced computational cost. The BasketLiDAR dataset contains a total of 4,445 frames and 3,105 player IDs, with fully synchronized IDs between three LiDAR sensors and three multi-view cameras. We recorded 5-on-5 and 3-on-3 game data from actual professional basketball players, providing complete 3D positional information and ID annotations for each player. Based on this dataset, we developed a novel MOT algorithm that leverages LiDAR's high-precision 3D spatial information. The proposed method consists of a real-time tracking pipeline using LiDAR alone and a multimodal tracking pipeline that fuses LiDAR and camera data. Experimental results demonstrate that our approach achieves real-time operation, which was difficult with conventional camera-only methods, while achieving superior tracking performance even under occlusion conditions. The dataset is available upon request at: https://sites.google.com/keio.jp/keio-csg/projects/basket-lidar.

Semi-supervised semantic segmentation (SS-SS) aims to mitigate the heavy annotation burden of dense pixel labeling by leveraging abundant unlabeled images alongside a small labeled set. While current consistency regularization methods achieve strong results, most do not explicitly model boundaries as a separate learning objective. In this paper, we propose BoundMatch, a novel multi-task SS-SS framework that explicitly integrates semantic boundary detection into a teacher-student consistency regularization pipeline. Our core mechanism, Boundary Consistency Regularized Multi-Task Learning (BCRM), enforces prediction agreement between teacher and student models on both segmentation masks and detailed semantic boundaries, providing complementary supervision from two independent tasks. To further enhance performance and encourage sharper boundaries, BoundMatch incorporates two lightweight fusion modules: Boundary-Semantic Fusion (BSF) injects learned boundary cues into the segmentation decoder, while Spatial Gradient Fusion (SGF) refines boundary predictions using mask gradients, yielding more reliable boundary pseudo-labels. This framework is built upon SAMTH, a strong teacher-student baseline featuring a Harmonious Batch Normalization (HBN) update strategy for improved stability. Extensive experiments on diverse datasets including Cityscapes and Pascal VOC show that BoundMatch achieves competitive performance against current state-of-the-art methods. Our approach achieves state-of-the-art results on the new Cityscapes benchmark with DINOv2 foundation model. Ablation studies highlight BoundMatch’s ability to improve boundary-specific evaluation metrics, its effectiveness in realistic large-scale unlabeled data scenario, and applicability to lightweight architectures for mobile deployment.

Background: Judges in figure skating are tasked with evaluating the technical quality of skaters’ performances using the grade of execution (GOE) score, a process that requires rapid integration of multiple technical and qualitative visual cues under time constraints, despite the inherently subjective nature of the evaluation criteria. Previous research suggests that judging accuracy is more impacted by previous experience as an athlete. However, the influence of judging experience on individual differences in evaluation remains unclear. Aims: This study investigated whether novice judging experience reduces individual differences in figure skating jump evaluations. Methods: Three certified judges and three skaters with similar athletic abilities evaluated 30 double Axel jumps from the 2019 World Championships using video-based judging. Participants assigned GOE scores and evaluated six GOE criteria while eye movements were recorded, and individual differences were analyzed using (1) judging accuracy, (2) gaze coordinates, and (3) intraclass correlations coefficients. Results and Conclusions: Results indicated that novice judging experience was not associated with improved judging accuracy or reduced individual differences in final GOE scores. However, judging experience was associated with lower dispersion in gaze coordinates and higher intraclass correlations coefficients in GOE criterion-based evaluations. Notably, substantial individual variation persisted when evaluating jumps matched to music, suggesting a challenge for judge development programs.

We propose an effective framework for the temporal action segmentation task, namely an Action Segment Refinement Framework (ASRF). Our model architecture consists of a long-term feature extractor and two branches: the Action Segmentation Branch (ASB) and the Boundary Regression Branch (BRB). The long-term feature extractor provides shared features for the two branches with a wide temporal receptive field. The ASB classifies video frames with action classes, while the BRB regresses the action boundary probabilities. The action boundaries predicted by the BRB refine the output from the ASB, which results in a significant performance improvement. Our contributions are three-fold: (i) We propose a framework for temporal action segmentation, the ASRF, which divides temporal action segmentation into frame-wise action classification and action boundary regression. Our framework refines frame-level hypotheses of action classes using predicted action boundaries. (ii) We propose a loss function for smoothing the transition of action probabilities, and analyze combinations of various loss functions for temporal action segmentation. (iii) Our framework outperforms state-of-the-art methods on three challenging datasets, offering an improvement of up to 13.7% in terms of segmental edit distance and up to 16.1% in terms of segmental F1 score. Our code is publicly available1.

Event cameras respond to scene dynamics and provide signals naturally suitable for motion estimation with advantages, such as high dynamic range. The emerging field of event-based vision motivates a revisit of fundamental computer vision tasks related to motion, such as optical flow and depth estimation. However, state-of-the-art event-based optical flow methods tend to originate in frame-based deep-learning methods, which require several adaptations (data conversion, loss function, etc.) as they have very different properties. We develop a principled method to extend the Contrast Maximization framework to estimate dense optical flow, depth, and ego-motion from events alone. The proposed method sensibly models the space-time properties of event data and tackles the event alignment problem. It designs the objective function to prevent overfitting, deals better with occlusions, and improves convergence using a multi-scale approach. With these key elements, our method ranks first among unsupervised methods on the MVSEC benchmark and is competitive on the DSEC benchmark. Moreover, it allows us to simultaneously estimate dense depth and ego-motion, exposes the limitations of current flow benchmarks, and produces remarkable results when it is transferred to unsupervised learning settings. Along with various downstream applications shown, we hope the proposed method becomes a cornerstone on event-based motion-related tasks. Code is available at https://github.com/tub-rip/event_based_optical_flow

Semantic segmentation is essential in computer vision for various applications, yet traditional approaches face significant challenges, including the high cost of annotation and extensive training for supervised learning. Additionally, due to the limited predefined categories in supervised learning, models typically struggle with infrequent classes and are unable to predict novel classes. To address these limitations, we propose MaskDiffusion, an innovative approach that leverages pretrained frozen Stable Diffusion to achieve open-vocabulary semantic segmentation without the need for additional training or annotation, leading to improved performance compared to similar methods. We also demonstrate the superior performance of MaskDiffusion in handling open vocabularies, including fine-grained and proper noun-based categories, thus expanding the scope of segmentation applications. Overall, our MaskDiffusion shows significant qualitative and quantitative improvements in contrast to other comparable unsupervised segmentation methods, i.e. on the Potsdam dataset (+10.5 mIoU compared to GEM) and COCO-Stuff (+14.8 mIoU compared to DiffSeg). All code and data are released at https://github.com/Valkyrja3607/MaskDiffusion.

Capturing the 3D human body is one of the important tasks in computer vision with a wide range of applications such as virtual reality and sports analysis. However, conventional frame cameras are limited by their temporal resolution and dynamic range, which imposes constraints in real-world application setups. Event cameras have the advantages of high temporal resolution and high dynamic range (HDR), but the development of event-based methods is necessary to handle data with different characteristics. This paper proposes a novel event-based method for 3D pose estimation and human mesh recovery. Prior work on event-based human mesh recovery require frames (images) as well as event data. The proposed method solely relies on events; it carves 3D voxels by moving the event camera around a stationary body, reconstructs the human pose and mesh by attenuated rays, and fit statistical body models, preserving high-frequency details. The experimental results show that the proposed method outperforms conventional frame-based methods in the estimation accuracy of both pose and body mesh. We also demonstrate results in challenging situations where other frame-based methods suffer from motion blur. This is the first-of-its-kind to demonstrate event-only human mesh recovery, and we hope that it is the first step toward achieving robust and accurate 3D human body scanning from vision sensors.

In recent years, advancements in the interaction and collaboration between humans and have garnered significant attention. Social intelligence plays a crucial role in facilitating natural interactions and seamless communication between humans and Artificial Intelligence (AI). To assess AI’s ability to understand human interactions and the components necessary for such comprehension, datasets like Social-IQ have been developed. However, these datasets often rely on a simplistic question-and-answer format and lack justifications for the provided answers. Furthermore, existing methods typically produce direct answers by selecting from predefined choices without generating intermediate outputs, which hampers interpretability and reliability. To address these limitations, we conducted a comprehensive evaluation of AI methods on a video-based Question Answering (QA) benchmark focused on human interactions, leveraging additional annotations related to human responses. Our analysis highlights significant differences between human and AI response patterns and underscores critical shortcomings in current benchmarks. We anticipate that these findings will guide the creation of more advanced datasets and represent an important step toward achieving natural communication between humans and AI.

Traditional Vision-and-Language Navigation (VLN) tasks require an agent to navigate static environments using natural language instructions. However, real-world road conditions such as vehicle movements, traffic signal fluctuations, pedestrian activity, and weather variations are dynamic and continually changing. These factors significantly impact an agent’s decision-making ability, underscoring the limitations of current VLN models, which do not accurately reflect the complexities of real-world navigation. To bridge this gap, we propose a novel task called Dynamic Vision-and-Language Navigation (DynamicVLN), incorporating various dynamic scenarios to enhance the agent’s decision-making abilities and adaptability. By redefining the VLN task, we emphasize that a robust and generalizable agent should not rely solely on predefined instructions but must also demonstrate reasoning skills and adaptability to unforeseen events. Specifically, we have designed ten scenarios that simulate the challenges of dynamic navigation and developed a dedicated dataset of 11,261 instances using the CARLA simulator (ver.0.9.13) and large language model to provide realistic training conditions. Additionally, we introduce a baseline model that integrates advanced perception and decision-making modules, enabling effective navigation and interpretation of the complexities of dynamic road conditions. This model showcases the ability to follow natural language instructions while dynamically adapting to environmental cues. Our approach establishes a benchmark for developing agents capable of functioning in real-world, dynamic environments and extending beyond the limitations of static VLN tasks to more practical and versatile applications.

Generating annotations for bird's-eye-view (BEV) segmentation presents significant challenges due to the scenes' complexity and the high manual annotation cost. In this work, we address these challenges by leveraging the abundance of unlabeled data available. We propose the Perspective Cue Training (PCT) framework, a novel training framework that utilizes pseudo-labels generated from unlabeled perspective images using publicly available semantic segmentation models trained on large street-view datasets. PCT applies a perspective view task head to the image encoder shared with the BEV segmentation head, effectively utilizing the unlabeled data to be trained with the generated pseudo-labels. Since image encoders are present in nearly all camera-based BEV segmentation architectures, PCT is flexible and applicable to various existing BEV architectures. In this paper, we applied PCT for semi-supervised learning (SSL) and unsupervised domain adaptation (UDA). Additionally, we introduce strong input perturbation through Camera Dropout (CamDrop) and feature perturbation via BEV Feature Dropout (BFD), which are crucial for enhancing SSL capabilities using our teacher-student framework. Our comprehensive approach is simple and flexible but yields significant improvements over various baselines for SSL and UDA, achieving competitive performances even against the current state-of-the-art.