Search

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.

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.

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.

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.

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.

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.

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 1.2% in IoU and a 2.6% 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.

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.

Abstract Background Knee osteoarthritis (OA) negatively affects dynamic postural control, which is a basic function that individuals use to perform activities of daily living (ADL). The purpose of this study was to investigate the associations of center of pressure (COP) control during the transition from double-leg to single-leg standing with subjective assessments of ADL and quality of life (QOL) in patients with knee OA. Methods Thirty-six patients (29 females) with moderate-to-severe knee OA participated. Dynamic postural control was evaluated during the transition from double-leg to single-leg standing. Each patient stood on a force plate, lifted the less affected limb as fast as possible, and maintained single-leg standing with the more affected limb. The COP movements corresponding to anticipatory postural adjustment (APA) and transitional phases were assessed. The maximum displacement and peak velocity of the COP movements in the medial–lateral direction were calculated. The Knee Injury and Osteoarthritis Outcome Score (KOOS) was used for the subjective assessment of ADL and QOL. Pearson’s product correlation analysis was performed to investigate the associations of COP movements in the APA and transitional phases with KOOS-ADL and KOOS-QOL. Results In the APA phase, the maximum COP displacement was significantly correlated with KOOS-ADL (r = -0.353, P = 0.035) and KOOS-QOL (r = -0.379, P = 0.023). In the transitional phase, the maximum COP displacement and peak COP velocity were significantly correlated with KOOS-ADL (maximum displacement: r = 0.352, P = 0.035; peak velocity: r = 0.438, P = 0.008) and with KOOS-QOL (maximum displacement: r = 0.357, P = 0.032; peak velocity: r = 0.343, P = 0.040). Conclusions The present study showed that smaller COP movements in the APA phase and smaller and slower COP movements in the transitional phase correlated with poorer ADL and QOL conditions in patients with knee OA. These findings suggest that poor dynamic postural control is associated with poor ADL and QOL conditions in patients with moderate-to-severe medial knee OA. Conservative treatment for patients with knee OA may need to focus on dynamic postural control during the transition from double-leg to single-leg standing.

Generative image extension has the advantage of extending the overall image size while preserving the target image because, unlike other image extensions using interpolation, it completes the surroundings of the target image. However, existing generative image extension methods tend to have poor quality in the generation of outer pixels. One method deals only with a limited number of scene classes because the extension repeats the same semantics. We propose a mirrored input, which sandwiches the extended region by mirroring a part of the target image. This replaces generative image extension with an image inpainting problem and thus helps to achieve higher quality pixel generation and can extend semantics with more complex shapes than horizontal repetition. Experimental results show that our proposed method achieves a scenery image extension that exceeds the state-of-the-art generative image extension methods in both visual quality and FID score for datasets containing diverse scenes.

The use of artificial intelligence (AI) as a tool supporting the diagnosis and treatment of spinal diseases is eagerly anticipated. In the field of diagnostic imaging, the possible application of AI includes diagnostic support for diseases requiring highly specialized expertise, such as trauma in children, scoliosis, symptomatic diseases, and spinal cord tumors. Moiré topography, which describes the 3-dimensional surface of the trunk with band patterns, has been used to screen students for scoliosis, but the interpretation of the band patterns can be ambiguous. Thus, we created a scoliosis screening system that estimates spinal alignment, the Cobb angle, and vertebral rotation from moiré images. In our system, a convolutional neural network (CNN) estimates the positions of 12 thoracic and 5 lumbar vertebrae, 17 spinous processes, and the vertebral rotation angle of each vertebra. We used this information to estimate the Cobb angle. The mean absolute error (MAE) of the estimated vertebral positions was 3.6 pixels (~5.4 mm) per person. T1 and L5 had smaller MAEs than the other levels. The MAE per person between the Cobb angle measured by doctors and the estimated Cobb angle was 3.42°. The MAE was 4.38° in normal spines, 3.13° in spines with a slight deformity, and 2.74° in spines with a mild to severe deformity. The MAE of the angle of vertebral rotation was 2.9°±1.4°, and was smaller when the deformity was milder. The proposed method of estimating the Cobb angle and AVR from moiré images using a CNN is expected to enhance the accuracy of scoliosis screening.

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.

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.

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 1.2 × 106 s to 8.3 × 104 s. The proposed method outperforms the current SoTA method by 1% accuracy on CIFAR-10.

Abstract Background A double-leg landing with or without a subsequent jump is commonly used to evaluate the neuromuscular control of knee abduction. However, the differences in frontal plane knee biomechanics between landings with and without a subsequent jump are not well known. The purpose of the present study was to investigate the effects of a subsequent jump on knee abduction, including during the early landing phase, in female and male subjects. Methods Twenty-one female subjects and 21 male subjects participated. All subjects performed drop landing task (a landing without a subsequent jump) and drop vertical jump task (a landing with a subsequent jump). The subjects landed from a 30-cm height. In drop vertical jump, the subjects also performed a maximum vertical jump immediately after landing. The knee abduction angle and moment were analyzed using a 3D motion analysis system. A two-way analysis of variance (task × time) was performed to examine the effects of a subsequent jump on the knee abduction angle during the early landing phase in female and male subjects. Another two-way analysis of variance (task × sex) was performed to compare peak knee abduction angles and moments. Results In female subjects, the knee abduction angle was significantly greater during drop vertical jump than during drop landing, as measured 45 to 80 ms after initial contact (P

Gaze estimation using monocular cameras has significant commercial applicability, and many studies have been undertaken on head pose-invariant and calibration-free gaze estimation. The head positions in existing data sets used in these studies are, however, limited to the vicinity of the camera, and methods trained on such data sets are not applicable when subjects are at greater distances from the camera. In this paper, we create a room-scale gaze data set with large variations in head poses to achieve robust gaze estimation across a broader range of widths and depths. The head positions are much farther from the camera, and the resolution of the eye image is lower than in conventional data sets. To address this issue, we propose a likelihood evaluation method based on edge gradients with dense particles for iris tracking, which achieves robust tracking at low-resolution eye images. Cross-validation experiments show that our proposed method is more accurate than conventional methods on all the individuals in our data set.

In this work, we propose a novel method of estimating optical flow from event-based cameras by matching the time surface of events. The proposed loss function measures the timestamp consistency between the time surface formed by the latest timestamp of each pixel and the one that is slightly shifted in time. This makes it possible to estimate dense optical flows with high accuracy without restoring luminance or additional sensor information. In the experiment, we show that the gradient was more correct and the loss landscape was more stable than the variance loss in the motion compensation approach. In addition, we show that the optical flow can be estimated with high accuracy by optimization with L1 smoothness regularization using publicly available datasets.

A hemi–closing-wedge and hemi–opening-wedge, inverted V–shaped high tibial osteotomy with local bone graft has been reported to be an effective surgical procedure for medial osteoarthritis of the knee. In this procedure, an inverted V–shaped osteotomy is made and a thin wedged bone block is resected from the lateral side and implanted in the medial opening space created after valgus correction. This procedure can provide sufficient valgus correction of the knee with severe varus deformity more easily than can closing-wedge high tibial osteotomy. The inverted V–shaped osteotomy does not change the posterior tibial slope, the patellar height, or the length of the lower limb at all because the center of tibial alignment correction by the inverted V–shaped osteotomy is located near the center of rotation of angulation of the lower-limb deformity. We recently modified this procedure by performing biplanar osteotomy, developing useful cutting guides, and fixing the tibia with a lateral locking compression plate. The surgical technique is described to enable the reproducible creation of the hemi–closing-wedge and hemi–opening-wedge, inverted V–shaped osteotomy with the locking plate for medial osteoarthritic knees with moderate or severe varus deformity.

We propose a novel method for seamline determination based on semantic segmentation for aerial image mosaicking. First, we train a convolutional neural network (CNN) for pixel labeling to extract building regions. Using the trained CNN, we create a building probability map from an input aerial image with no pre-processing. We then use Dijkstra's algorithm to find the optimal seamline as a shortest path on the map. We evaluate the quality of the seamlines produced by our method on actual aerial images. Finally, we show that our seamlines never pass through any buildings and compare the effectiveness with the conventional mean-shift segmentation-based method.

The paper presents an emerging issue of fine-grained pedestrian action recognition that induces an advanced pre-crush safety to estimate a pedestrian intention in advance. The fine-grained pedestrian actions include visually slight differences (e.g., walking straight and crossing), which are difficult to distinguish from each other. It is believed that the fine-grained action recognition induces a pedestrian intention estimation for a helpful advanced driver-assistance systems (ADAS). The following difficulties have been studied to achieve a fine-grained and accurate pedestrian action recognition: (i) In order to analyze the fine-grained motion of a pedestrian appearance in the vehicle-mounted drive recorder, a method to describe subtle change of motion characteristics occurring in a short time is necessary; (ii) even when the background moves greatly due to the driving of the vehicle, it is necessary to detect changes in subtle motion of the pedestrian; (iii) the collection of large-scale fine-grained actions is very difficult, and therefore a relatively small database should be focused. We find out how to learn an effective recognition model with only a small-scale database. Here, we have thoroughly evaluated several types of configurations to explore an effective approach in fine-grained pedestrian action recognition without a large-scale database. Moreover, two different datasets have been collected in order to raise the issue. Finally, our proposal attained 91.01% on National Traffic Science and Environment Laboratory database (NTSEL) and 53.23% on the near-miss driving recorder database (NDRDB). The paper has improved +8.28% and +6.53% from baseline two-stream fusion convnets.