論文

In online video understanding, which has a wide range of real-world applications, inference speed is crucial. Many approaches involve frame-level visual feature extraction, which often represents the biggest bottleneck. We propose RetinaViT, an efficient method for extracting frame-level visual features in an online video stream, aiming to fundamentally enhance the efficiency of online video understanding tasks. RetinaViT is composed of efficiently approximated Transformer blocks that only take changed tokens (event tokens) as queries and reuse the already processed tokens from the previous timestep for the others. Furthermore, we restrict keys and values to the spatial neighborhoods of event tokens to further improve efficiency. RetinaViT involves tuning multiple parameters, which we determine through a multi-step process. During model training, we randomly vary these parameters and then perform black-box optimization to maximize accuracy and efficiency on the pre-trained model. We conducted extensive experiments on various online video recognition tasks, including action recognition, pose estimation, and object segmentation, validating the effectiveness of each component in RetinaViT and demonstrating improvements in the speed/accuracy trade-off compared to baselines. In particular, for action recognition, RetinaViT built on ViT-B16 reduces inference time by approximately 61.9% on the CPU and 50.8% on the GPU, while achieving slight accuracy improvements rather than degradation.

Semantic segmentation is a crucial task in computer vision, where each pixel in an image is classified into a category. However, traditional methods face significant challenges, including the need for pixel-level annotations and extensive training. Furthermore, because supervised learning uses a limited set of predefined categories, models typically struggle with rare classes and cannot recognize new ones. Unsupervised and open-vocabulary segmentation, proposed to tackle these issues, faces challenges, including the inability to assign specific class labels to clusters and the necessity of user-provided text queries for guidance. In this context, we propose a novel approach, TAG which achieves Training, Annotation, and Guidance-free open-vocabulary semantic segmentation. TAG utilizes pre-trained models such as CLIP and DINO to segment images into meaningful categories without additional training or dense annotations. It retrieves class labels from an external database, providing flexibility to adapt to new scenarios. Our TAG achieves state-of-the-art results on PascalVOC, PascalContext and ADE20K for open-vocabulary segmentation without given class names, i.e. improvement of +15.3 mIoU on PascalVOC.

We address the problem of generating a 360-degree image from a single image with a narrow field of view by estimating its surroundings. Previous methods suffered from overfitting to the training resolution and deterministic generation. This paper proposes a completion method using a transformer for scene modeling and novel methods to improve the properties of a 360-degree image on the output image. Specifically, we use CompletionNets with a transformer to perform diverse completions and Adjust-mentNet to match color, stitching, and resolution with an input image, enabling inference at any resolution. To improve the properties of a 360-degree image on an output image, we also propose WS-perceptual loss and circular inference. Thorough experiments show that our method out-performs state-of-the-art (SOTA) methods both qualitatively and quantitatively. For example, compared to SOTA methods, our method completes images 16 times larger in resolution and achieves 1.7 times lower Fréchet inception distance (FID). Furthermore, we propose a pipeline that uses the completion results for lighting and background of 3DCG scenes. Our plausible background completion enables perceptually natural results in the application of inserting virtual objects with specular surfaces.

Given only acoustic signals without any high-level information, such as voices or sounds of scenes/actions, how much can we infer about the behavior of humans? Unlike existing methods, which suffer from privacy issues because they use signals that include human speech or the sounds of specific actions, we explore how low-level acoustic signals can provide enough clues to estimate 3D human poses by active acoustic sensing with a single pair of microphones and loudspeakers (see Fig. 1). This is a challenging task since sound is much more diffractive than other signals and therefore covers up the shape of objects in a scene. Accordingly, we introduce a framework that encodes multichannel audio features into 3D human poses. Aiming to capture subtle sound changes to reveal detailed pose information, we explicitly extract phase features from the acoustic signals together with typical spectrum features and feed them into our human pose estimation network. Also, we show that reflected or diffracted sounds are easily influenced by subjects' physique differences e.g., height and muscularity, which deteriorates prediction accuracy. We reduce these gaps by using a subject discriminator to improve accuracy. Our experiments suggest that with the use of only low-dimensional acoustic information, our method outperforms baseline methods. The datasets and codes used in this project will be publicly available.

Large vision-language models, such as Contrastive Vision-Language Pre-training (CLIP), pre-trained on large-scale image–text datasets, have demonstrated robust zero-shot transfer capabilities across various downstream tasks. To further enhance the few-shot recognition performance of CLIP, Tip-Adapter augments the CLIP model with an adapter that incorporates a key-value cache model constructed from the few-shot training set. This approach enables training-free adaptation and has shown significant improvements in few-shot recognition, especially with additional fine-tuning. However, the size of the adapter increases in proportion to the number of training samples, making it difficult to deploy in practical applications. In this paper, we propose a novel CLIP adaptation method, named Proto-Adapter, which employs a single-layer adapter of constant size regardless of the amount of training data and even outperforms Tip-Adapter. Proto-Adapter constructs the adapter’s weights based on prototype representations for each class. By aggregating the features of the training samples, it successfully reduces the size of the adapter without compromising performance. Moreover, the performance of the model can be further enhanced by fine-tuning the adapter’s weights using a distance margin penalty, which imposes additional inter-class discrepancy to the output logits. We posit that this training scheme allows us to obtain a model with a discriminative decision boundary even when trained with a limited amount of data. We demonstrate the effectiveness of the proposed method through extensive experiments of few-shot classification on diverse datasets.

Schlieren imaging is an optical technique to observe the flow of transparent media, such as air or water, without any particle seeding. However, conventional frame-based techniques require both high spatial and temporal resolution cameras, which impose bright illumination and expensive computation limitations. Event cameras offer potential advantages (high dynamic range, high temporal resolution, and data efficiency) to overcome such limitations due to their bio-inspired sensing principle. This article presents a novel technique for perceiving air convection using events and frames by providing the first theoretical analysis that connects event data and schlieren. We formulate the problem as a variational optimization one combining the linearized event generation model with a physically-motivated parameterization that estimates the temporal derivative of the air density. The experiments with accurately aligned frame- and event camera data reveal that the proposed method enables event cameras to obtain on par results with existing frame-based optical flow techniques. Moreover, the proposed method works under dark conditions where frame-based schlieren fails, and also enables slow-motion analysis by leveraging the event camera's advantages. Our work pioneers and opens a new stack of event camera applications, as we publish the source code as well as the first schlieren dataset with high-quality frame and event data.

Action quality assessment (AQA) tasks in computer vision evaluate action quality in videos, and they can be applied to sports for performance evaluation. A typical example of AQA is predicting the final score from a video that captures an entire figure skating program. However, no previous studies have predicted individual jump scores, which are of great interest to competitors because of the high weight of competition. Despite the presence of unnecessary information in figure skating videos, human specialists can focus and reduce information when they evaluate jumps. In this study, we clarified the eye movements of figure skating judges and skaters while evaluating jumps and proposed a prediction model for jump performance that utilized specialists’ gaze location to reduce information. Kinematic features obtained from the tracking system were input into the model in addition to videos to improve accuracy. The results showed that skaters focused more on the face, whereas judges focused on the lower extremities. These gaze locations were applied to the model, which demonstrated the highest accuracy when utilizing both specialists’ gaze locations. The model outperformed human predictions and the baseline model (RMSE:0.775), suggesting a combination of human specialist knowledge and machine capabilities could yield higher accuracy.

Shadow removal for document images is an essential 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, diverse dataset for document shadow removal takes time and effort. Thus, only small real datasets are available. Graphic renderers have been used to synthesize shadows to create relatively large datasets. However, the limited number of unique documents and the limited lighting environments adversely affect the network performance. This paper presents a large-scale, diverse dataset called the Synthetic Document with Diverse Shadows (SynDocDS) dataset. The SynDocDS comprises rendered images with diverse shadows augmented by a physics-based illumination model, which can be utilized to obtain a more robust and high-performance deep shadow removal network. In this paper, we further propose a Dual Shadow Fusion Network (DSFN). Unlike natural images, document images often have constant background colors requiring a high understanding of global color features for training a deep shadow removal network. The DSFN has a high global color comprehension and understanding of shadow regions and merges shadow attentions and features efficiently. We conduct experiments on three publicly available datasets, the OSR, Kligler’s, and Jung’s datasets, to validate our proposed method’s effectiveness. In comparison to training on existing synthetic datasets, our model training on the SynDocDS dataset achieves an enhancement in the PSNR and SSIM, increasing them from 23.00 dB to 25.70 dB and 0.959 to 0.971 on average. In addition, the experiments demonstrated that our DSFN clearly outperformed other networks across multiple metrics, including the PSNR, the SSIM, and its impact on OCR performance.

Vision-based human activity classification has remarkable potential for various applications in the sports context (e.g., motion analysis for performance enhancement, active sensing for athletes, etc.). Recently, learning-based human activity classifications have been widely researched. However, in sports scenes in which more detailed and player-specific classifications are required, this is a quite challenging task; in many cases, only a limited number of datasets are available, unlike daily movements such as walking or climbing stairs. Therefore, this paper proposes a time-weighted motion history image, an effective image sequence representation for learning-based human activity classification. Unlike conventional MHI based on the assumption that “the newer frame is more important,” our method generates importance-aware representation so that the predictor can “see” the frames that contribute to analyzing the specific human activity. Experimental results have shown the superiority of our method.

Event cameras are emerging vision sensors and their advantages are suitable for various applications such as autonomous robots. Contrast maximization (CMax), which provides state-of-the-art accuracy on motion estimation using events, may suffer from an overfitting problem called event collapse. Prior works are computationally expensive or cannot alleviate the overfitting, which undermines the benefits of the CMax framework. A novel, computationally efficient regularizer based on geometric principles to mitigate event collapse is proposed. The experiments show that the proposed regularizer achieves state-of-the-art accuracy results, while its reduced computational complexity makes it two to four times faster than previous approaches. To the best of our knowledge, this regularizer is the only effective solution for event collapse without trading off the runtime. It is hoped that this work opens the door for future applications that unlocks the advantages of event cameras. Project page: https://github.com/tub-rip/event_collapse.

In this paper, we propose the Semantic-Boundary-Conditioned Backbone (SBCB) framework, an effective approach to enhancing semantic segmentation performance, particularly around mask boundaries, while maintaining compatibility with various segmentation architectures. Our objective is to improve existing models by leveraging semantic boundary information as an auxiliary task. The SBCB framework incorporates a complementary semantic boundary detection (SBD) task with a multi-task learning approach. It enhances the segmentation backbone without introducing additional parameters during inference or relying on independent post-processing modules. The SBD head utilizes multi-scale features from the backbone, learning low-level features in early stages and understanding high-level semantics in later stages. This complements common semantic segmentation architectures, where features from later stages are used for classification. Extensive evaluations using popular segmentation heads and backbones demonstrate the effectiveness of the SBCB. It leads to an average improvement of (Formula presented.) in IoU and a (Formula presented.) gain in the boundary F-score on the Cityscapes dataset. The SBCB framework also improves over- and under-segmentation characteristics. Furthermore, the SBCB adapts well to customized backbones and emerging vision transformer models, consistently achieving superior performance. In summary, the SBCB framework significantly boosts segmentation performance, especially around boundaries, without introducing complexity to the models. Leveraging the SBD task as an auxiliary objective, our approach demonstrates consistent improvements on various benchmarks, confirming its potential for advancing the field of semantic segmentation.

Recently, Transformer-based video recognition models have achieved state-of-the-art results on major video recognition benchmarks. However, their high inference cost significantly limits research speed and practical use. In video compression, methods considering small motions and residuals that are less informative and assigning short code lengths to them (e.g., MPEG4) have successfully reduced the redundancy of videos. Inspired by this idea, we propose Informative Patch Selection (IPS), which efficiently reduces the inference cost by excluding redundant patches from the input of the Transformer-based video model. The redundancy of each patch is calculated from motions and residuals obtained while decoding a compressed video. The proposed method is simple and effective in that it can dynamically reduce the inference cost depending on the input without any policy model or additional loss term. Extensive experiments on action recognition demonstrated that our method could significantly improve the trade-off between the accuracy and inference cost of the Transformer-based video model. Although the method does not require any policy model or additional loss term, its performance approaches that of existing methods that do require them.

Event cameras are novel bio-inspired sensors that offer advantages over traditional cameras (low latency, high dynamic range, low power, etc.). Optical flow estimation methods that work on packets of events trade off speed for accuracy, while event-by-event (incremental) methods have strong assumptions and have not been tested on common benchmarks that quantify progress in the field. Towards applications on resource-constrained devices, it is important to develop optical flow algorithms that are fast, light-weight and accurate. This work leverages insights from neuroscience, and proposes a novel optical flow estimation scheme based on triplet matching. The experiments on publicly available benchmarks demonstrate its capability to handle complex scenes with comparable results as prior packet-based algorithms. In addition, the proposed method achieves the fastest execution time (>10 kHz) on standard CPUs as it requires only three events in estimation. We hope that our research opens the door to real-time, incremental motion estimation methods and applications in real-world scenarios.

In the field of embodied AI, vision-and-language navigation (VLN) is a crucial and challenging multi-modal task. Specifically, outdoor VLN involves an agent navigating within a graph-based environment, while simultaneously interpreting information from real-world urban environments and natural language instructions. Existing outdoor VLN models predict actions using a combination of panorama and instruction features. However, these methods may cause the agent to struggle to understand complicated outdoor environments and ignore the details in the environments to fail to navigate. Human navigation often involves the use of specific objects as reference landmarks when navigating to unfamiliar places, providing a more rational and efficient approach to navigation. Inspired by this natural human behavior, we propose an object-level alignment module (OAlM), which guides the agent to focus more on object tokens mentioned in the instructions and recognize these landmarks during navigation. By treating these landmarks as sub-goals, our method effectively decomposes a long-range path into a series of shorter paths, ultimately improving the agent’s overall performance. In addition to enabling better object recognition and alignment, our proposed OAlM also fosters a more robust and adaptable agent capable of navigating complex environments. This adaptability is particularly crucial for real-world applications where environmental conditions can be unpredictable and varied. Experimental results show our OAlM is a more object-focused model, and our approach outperforms all metrics on a challenging outdoor VLN Touchdown dataset, exceeding the baseline by 3.19% on task completion (TC). These results highlight the potential of leveraging object-level information in the form of sub-goals to improve navigation performance in embodied AI systems, paving the way for more advanced and efficient outdoor navigation.

One way to improve annotation efficiency is active learning. The goal of active learning is to select images from many unlabeled images, where labeling will improve the accuracy of the machine learning model the most. To select the most informative unlabeled images, conventional methods use deep neural networks with a large number of computation nodes and long computation time, but we propose a non-deep neural network method that does not require any additional training for unlabeled image selection. The proposed method trains a task model on labeled images, and then the model predicts unlabeled images. Based on this prediction, an uncertainty indicator is generated for each unlabeled image. Images with a high uncertainty index are considered to have a high information content, and are selected for annotation. Our proposed method is based on a very simple and powerful idea: select samples near the decision boundary of the model. Experimental results on multiple datasets show that the proposed method achieves higher accuracy than conventional active learning methods on multiple tasks and up to 14 times faster execution time from (Formula presented.) s to (Formula presented.) s. The proposed method outperforms the current SoTA method by 1% accuracy on CIFAR-10.

One of the major factors that makes the task of tracking multiple objects difficult is frequent occlusions in congestion situations. Many existing methods in multi-object tracking by online processing have adopted a tracking-by-detection approach to perform object detection in every frame of a video and associate the object bounding box obtained by it temporally. However, with the existing method, it was not possible to track objects that could not be acquired by detectors due to occlusion. Therefore, we propose a method to shift the state of the lost target to tracked state once again by tracklet re-identification. Embeddings expressing the high dimensional appearance features of a object are acquired by using a convolutional neural network and re-identification of the tracklet is made based on the distance of the embedding vector among the tracklets. At this time, by using the masked image of the object obtained by instance segmentation as the input of the network, it is possible to perform re-identification determination robust to the background change. Further, since the re-identification determination of the tracklet pair is performed based on the distance between the low-dimensional vectors, the increase in the calculation cost due to the re-identification processing is small. Experiments were carried out using MOT16 dataset, which is a public data set, and the effectiveness of this method is shown.

A figure skating jump score is determined by the sum of the base value based on the difficulty and grade of execution (GOE) that indicates the performance quality. Therefore, performing a high-quality jump to obtain a high GOE is essential to win a competition. However, the relationship between the GOE and kinematic parameters remains unclear. We analysed the horizontal distance, vertical height, and landing speed of double axel jumps in the Ladies’ Short Program at the 2019 World Championships. The highest GOE group had significantly larger horizontal distances than the middle and lower groups, while the landing speed and vertical height were not significantly different. A principal component regression analysis was conducted to clarify the contrast between the three variables affecting the GOE. The results showed that greater horizontal distance and landing speed compared to vertical height (component 1) and greater horizontal distance compared to landing speed (component 3) contributed to higher GOE. We divided skaters into four clusters using these two components and provided general GOE acquisition strategies for each cluster. Finally, to apply our results to the industry, we proposed two new evaluation indicators which are highly correlated with the two components and easy to interpret.

Contrast maximization (CMax) is a framework that provides state-of-the-art results on several event-based computer vision tasks, such as ego-motion or optical flow estimation. However, it may suffer from a problem called event collapse, which is an undesired solution where events are warped into too few pixels. As prior works have largely ignored the issue or proposed workarounds, it is imperative to analyze this phenomenon in detail. Our work demonstrates event collapse in its simplest form and proposes collapse metrics by using first principles of space–time deformation based on differential geometry and physics. We experimentally show on publicly available datasets that the proposed metrics mitigate event collapse and do not harm well-posed warps. To the best of our knowledge, regularizers based on the proposed metrics are the only effective solution against event collapse in the experimental settings considered, compared with other methods. We hope that this work inspires further research to tackle more complex warp models.

© 2018, Springer International Publishing AG, part of Springer Nature. In this paper we propose a novel approach for video-based person re-identification that exploits convolutional neural networks to learn the similarity of persons observed from video camera. We take 3-dimensional convolutional neural networks (3D CNN) to extract fine-grained spatiotemporal features from the video sequence of a person. Unlike recurrent neural networks, 3D CNN preserves the spatial patterns of the input, which works well on re-identification problem. The network maps each video sequence of a person to a Euclidean space where distances between feature embeddings directly correspond to measures of person similarity. By our improved parameter learning method called entire triplet loss, all possible triplets in the mini-batch are taken into account to update network parameters. This parameter updating method significantly improves training, enabling the embeddings to be more discriminative. Experimental results show that our model achieves new state of the art identification rate on iLIDS-VID dataset and PRID-2011 dataset with 82.0%, 83.3% at rank 1, respectively.

In this paper, a novel setting is tackled in which a neural network generates object images with transferred attributes, by conditioning on natural language commands. Conventional methods for object image transformation have used visual attributes, which are components that describe the object's color, posture, etc. This paper builds on this approach and finds an algorithm to precisely extract information from natural language commands, which transfers the attributes of an image and completes this image translation model. The effectiveness of our information extraction model is experimented, with additional tests to see if the change in visual attributes is correctly seen in the image.