Search

Image-based Pose-Agnostic 3D Anomaly Detection is an important task that has emerged in industrial quality control. This task seeks to find anomalies from query images of a tested object given a set of reference images of an anomaly-free object. The challenge is that the query views (a.k.a poses) are unknown and can be different from the reference views. Currently, new methods such as OmniposeAD and SplatPose have emerged to bridge the gap by synthesizing pseudo reference images at the query views for pixel-to-pixel comparison. However, none of these methods can infer in real-time, which is critical in industrial quality control for massive production. For this reason, we propose SplatPose+, which employs a hybrid representation consisting of a Structure from Motion (SfM) model for localization and a 3D Gaussian Splatting (3DGS) model for Novel View Synthesis. Although our proposed pipeline requires the computation of an additional SfM model, it offers real-time inference speeds and faster training compared to SplatPose. Quality-wise, we achieved a new SOTA on the Pose-agnostic Anomaly Detection benchmark with the Multi-Pose Anomaly Detection (MAD-SIM) dataset.

This paper presents an integrated approach to Visual SLAM, merging online sequential photometric calibration within a Hybrid direct-indirect visual SLAM (H-SLAM). Photometric calibration helps normalize pixel intensity values under different lighting conditions, and thereby improves the direct component of our H-SLAM. A tangential benefit also results to the indirect component of H-SLAM given that the detected features are more stable across variable lighting conditions. Our proposed photometrically calibrated H-SLAM is tested on several datasets, including the TUM monoVO as well as on a dataset we created. Calibrated H-SLAM outperforms other state of the art direct, indirect, and hybrid Visual SLAM systems in all the experiments. Furthermore, in online SLAM tested at our site, it also significantly outperformed the other SLAM Systems.

Real-time SLAM with dense 3D mapping is computationally challenging, especially on resource-limited devices. The recent development of 3D Gaussian Splatting (3DGS) offers a promising approach for real-time dense 3D reconstruction. However, existing 3DGS-based SLAM systems struggle to balance hardware simplicity, speed, and map quality. Most systems excel in one or two of the aforementioned aspects but rarely achieve all. A key issue is the difficulty of initializing 3D Gaussians while concurrently conducting SLAM. To address these challenges, we present Monocular GSO (MGSO), a novel real-time SLAM system that integrates photometric SLAM with 3DGS. Photometric SLAM provides dense structured point clouds for 3DGS initialization, accelerating optimization and producing more efficient maps with fewer Gaussians. As a result, experiments show that our system generates reconstructions with a balance of quality, memory efficiency, and speed that outperforms the state-of-the-art. Furthermore, our system achieves all results using RGB inputs. We evaluate the Replica, TUM-RGBD, and EuRoC datasets against current live dense reconstruction systems. Not only do we surpass contemporary systems, but experiments also show that we maintain our performance on laptop hardware, making it a practical solution for robotics, A/R, and other real-time applications., Comment: Paper Contribution to the ICRA 2025 Conference. Currently being reviewed

The application of autonomous robotics to close-contact healthcare tasks has a clear role for the future due to its potential to reduce infection risks to staff and improve clinical efficiency. Nasopharyngeal (NP) swab sample collection for diagnosing upper-respiratory illnesses is one type of close contact task that is interesting for robotics due to the dexterity requirements and the unobservability of the nasal cavity. We propose a control system that performs the test using a collaborative manipulator arm with an instrumented end-effector to take visual and force measurements, under the scenario that the patient is unrestrained and the tools are general enough to be applied to other close contact tasks. The system employs a visual servo controller to align the swab with the nostrils. A compliant joint velocity controller inserts the swab along a trajectory optimized through a simulation environment, that also reacts to measured forces applied to the swab. Additional subsystems include a fuzzy logic system for detecting when the swab reaches the nasopharynx and a method for detaching the swab and aborting the procedure if safety criteria is violated. The system is evaluated using a second robotic arm that holds a nasal cavity phantom and simulates the natural head motions that could occur during the procedure. Through extensive experiments, we identify controller configurations capable of effectively performing the NP swab test even with significant head motion, which demonstrates the safety and reliability of the system., Comment: 18 pages, 26 figures. Supplementary files may be found at https://uwaterloo.ca/scholar/pqjlee/supplementary-files-complete-autonomous-robotic-nasopharyngeal-swab-system-evaluation

The nasopharyngeal (NP) swab test is a method for collecting cultures to diagnose for different types of respiratory illnesses, including COVID-19. Delegating this task to robots would be beneficial in terms of reducing infection risks and bolstering the healthcare system, but a critical component of the NP swab test is having the swab aligned properly with the nasal cavity so that it does not cause excessive discomfort or injury by traveling down the wrong passage. Existing research towards robotic NP swabbing typically assumes the patient's head is held within a fixture. This simplifies the alignment problem, but is also dissimilar to clinical scenarios where patients are typically free-standing. Consequently, our work creates a vision-guided pipeline to allow an instrumented robot arm to properly position and orient NP swabs with respect to the nostrils of free-standing patients. The first component of the pipeline is a precomputed joint lookup table to allow the arm to meet the patient's arbitrary position in the designated workspace, while avoiding joint limits. Our pipeline leverages semantic face models from computer vision to estimate the Euclidean pose of the face with respect to a monocular RGB-D camera placed on the end-effector. These estimates are passed into an unscented Kalman filter on manifolds state estimator and a pose based visual servo control loop to move the swab to the designated pose in front of the nostril. Our pipeline was validated with human trials, featuring a cohort of 25 participants. The system is effective, reaching the nostril for 84% of participants, and our statistical analysis did not find significant demographic biases within the cohort., Comment: 12 pages, 13 figures

The nasopharyngeal (NP) swab sample test, commonly used to detect COVID-19 and other respiratory illnesses, involves moving a swab through the nasal cavity to collect samples from the nasopharynx. While typically this is done by human healthcare workers, there is a significant societal interest to enable robots to do this test to reduce exposure to patients and to free up human resources. The task is challenging from the robotics perspective because of the dexterity and safety requirements. While other works have implemented specific hardware solutions, our research differentiates itself by using a ubiquitous rigid robotic arm. This work presents a case study where we investigate the strengths and challenges using compliant control system to accomplish NP swab tests with such a robotic configuration. To accomplish this, we designed a force sensing end-effector that integrates with the proposed torque controlled compliant control loop. We then conducted experiments where the robot inserted NP swabs into a 3D printed nasal cavity phantom. Ultimately, we found that the compliant control system outperformed a basic position controller and shows promise for human use. However, further efforts are needed to ensure the initial alignment with the nostril and to address head motion., Comment: 13 pages, 9 figures. See https://uwaterloo.ca/scholar/pqjlee/collaborative-robot-arm-inserting-nasopharyngeal-swabs-admittance-control for supplementary data

Initial applications of 3D Gaussian Splatting (3DGS) in Visual Simultaneous Localization and Mapping (VSLAM) demonstrate the generation of high-quality volumetric reconstructions from monocular video streams. However, despite these promising advancements, current 3DGS integrations have reduced tracking performance and lower operating speeds compared to traditional VSLAM. To address these issues, we propose integrating 3DGS with Direct Sparse Odometry, a monocular photometric SLAM system. We have done preliminary experiments showing that using Direct Sparse Odometry point cloud outputs, as opposed to standard structure-from-motion methods, significantly shortens the training time needed to achieve high-quality renders. Reducing 3DGS training time enables the development of 3DGS-integrated SLAM systems that operate in real-time on mobile hardware. These promising initial findings suggest further exploration is warranted in combining traditional VSLAM systems with 3DGS., Comment: This extended abstract has been submitted to be presented at an IEEE conference. It will be made available online by IEEE but will not be published in IEEE Xplore

Motion segmentation from a single moving camera presents a significant challenge in the field of computer vision. This challenge is compounded by the unknown camera movements and the lack of depth information of the scene. While deep learning has shown impressive capabilities in addressing these issues, supervised models require extensive training on massive annotated datasets, and unsupervised models also require training on large volumes of unannotated data, presenting significant barriers for both. In contrast, traditional methods based on optical flow do not require training data, however, they often fail to capture object-level information, leading to over-segmentation or under-segmentation. In addition, they also struggle in complex scenes with substantial depth variations and non-rigid motion, due to the overreliance of optical flow. To overcome these challenges, we propose an innovative hybrid approach that leverages the advantages of both deep learning methods and traditional optical flow based methods to perform dense motion segmentation without requiring any training. Our method initiates by automatically generating object proposals for each frame using foundation models. These proposals are then clustered into distinct motion groups using both optical flow and relative depth maps as motion cues. The integration of depth maps derived from state-of-the-art monocular depth estimation models significantly enhances the motion cues provided by optical flow, particularly in handling motion parallax issues. Our method is evaluated on the DAVIS-Moving and YTVOS-Moving datasets, and the results demonstrate that our method outperforms the best unsupervised method and closely matches with the state-of-theart supervised methods., Comment: For the offical publication, see https://crv.pubpub.org/pub/iunjzl55

Novel dataset creation for all multi-object tracking, crowd-counting, and industrial-based videos is arduous and time-consuming when faced with a unique class that densely populates a video sequence. We propose a time efficient method called POPCat that exploits the multi-target and temporal features of video data to produce a semi-supervised pipeline for segmentation or box-based video annotation. The method retains the accuracy level associated with human level annotation while generating a large volume of semi-supervised annotations for greater generalization. The method capitalizes on temporal features through the use of a particle tracker to expand the domain of human-provided target points. This is done through the use of a particle tracker to reassociate the initial points to a set of images that follow the labeled frame. A YOLO model is then trained with this generated data, and then rapidly infers on the target video. Evaluations are conducted on GMOT-40, AnimalTrack, and Visdrone-2019 benchmarks. These multi-target video tracking/detection sets contain multiple similar-looking targets, camera movements, and other features that would commonly be seen in "wild" situations. We specifically choose these difficult datasets to demonstrate the efficacy of the pipeline and for comparison purposes. The method applied on GMOT-40, AnimalTrack, and Visdrone shows a margin of improvement on recall/mAP50/mAP over the best results by a value of 24.5%/9.6%/4.8%, -/43.1%/27.8%, and 7.5%/9.4%/7.5% where metrics were collected., Comment: 10 pages, 5 figures, Accepted in "Conference on Robots and Vision 2024"

Multi Object Tracking (MOT) in ice hockey pursues the combined task of localizing and associating players across a given sequence to maintain their identities. Tracking players from monocular broadcast feeds is an important computer vision problem offering various downstream analytics and enhanced viewership experience. However, existing trackers encounter significant difficulties in dealing with occlusions, blurs, and agile player movements prevalent in telecast feeds. In this work, we propose a novel tracking approach by formulating MOT as a bipartite graph matching problem infused with homography. We disentangle the positional representations of occluded and overlapping players in broadcast view, by mapping their foot keypoints to an overhead rink template, and encode these projected positions into the graph network. This ensures reliable spatial context for consistent player tracking and unfragmented tracklet prediction. Our results show considerable improvements in both the IDsw and IDF1 metrics on the two available broadcast ice hockey datasets., Comment: Accepted at the Conference on Robots and Vision (CRV), 2024

In the high-stakes world of baseball, every nuance of a pitcher's mechanics holds the key to maximizing performance and minimizing runs. Traditional analysis methods often rely on pre-recorded offline numerical data, hindering their application in the dynamic environment of live games. Broadcast video analysis, while seemingly ideal, faces significant challenges due to factors like motion blur and low resolution. To address these challenges, we introduce PitcherNet, an end-to-end automated system that analyzes pitcher kinematics directly from live broadcast video, thereby extracting valuable pitch statistics including velocity, release point, pitch position, and release extension. This system leverages three key components: (1) Player tracking and identification by decoupling actions from player kinematics; (2) Distribution and depth-aware 3D human modeling; and (3) Kinematic-driven pitch statistics. Experimental validation demonstrates that PitcherNet achieves robust analysis results with 96.82% accuracy in pitcher tracklet identification, reduced joint position error by 1.8mm and superior analytics compared to baseline methods. By enabling performance-critical kinematic analysis from broadcast video, PitcherNet paves the way for the future of baseball analytics by optimizing pitching strategies, preventing injuries, and unlocking a deeper understanding of pitcher mechanics, forever transforming the game., Comment: IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW'24)

Detecting and segmenting moving objects from a moving monocular camera is challenging in the presence of unknown camera motion, diverse object motions and complex scene structures. Most existing methods rely on a single motion cue to perform motion segmentation, which is usually insufficient when facing different complex environments. While a few recent deep learning based methods are able to combine multiple motion cues to achieve improved accuracy, they depend heavily on vast datasets and extensive annotations, making them less adaptable to new scenarios. To address these limitations, we propose a novel monocular dense segmentation method that achieves state-of-the-art motion segmentation results in a zero-shot manner. The proposed method synergestically combines the strengths of deep learning and geometric model fusion methods by performing geometric model fusion on object proposals. Experiments show that our method achieves competitive results on several motion segmentation datasets and even surpasses some state-of-the-art supervised methods on certain benchmarks, while not being trained on any data. We also present an ablation study to show the effectiveness of combining different geometric models together for motion segmentation, highlighting the value of our geometric model fusion strategy., Comment: Accepted by the 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW)

Unique player identification is a fundamental module in vision-driven sports analytics. Identifying players from broadcast videos can aid with various downstream tasks such as player assessment, in-game analysis, and broadcast production. However, automatic detection of jersey numbers using deep features is challenging primarily due to: a) motion blur, b) low resolution video feed, and c) occlusions. With their recent success in various vision tasks, masked autoencoders (MAEs) have emerged as a superior alternative to conventional feature extractors. However, most MAEs simply zero-out image patches either randomly or focus on where to mask rather than how to mask. Motivated by human vision, we devise a novel domain-guided masking policy for MAEs termed d-MAE to facilitate robust feature extraction in the presence of motion blur for player identification. We further introduce a new spatio-temporal network leveraging our novel d-MAE for unique player identification. We conduct experiments on three large-scale sports datasets, including a curated baseball dataset, the SoccerNet dataset, and an in-house ice hockey dataset. We preprocess the datasets using an upgraded keyframe identification (KfID) module by focusing on frames containing jersey numbers. Additionally, we propose a keyframe-fusion technique to augment keyframes, preserving spatial and temporal context. Our spatio-temporal network showcases significant improvements, surpassing the current state-of-the-art by 8.58%, 4.29%, and 1.20% in the test set accuracies, respectively. Rigorous ablations highlight the effectiveness of our domain-guided masking approach and the refined KfID module, resulting in performance enhancements of 1.48% and 1.84% respectively, compared to original architectures., Comment: Submitted to 21st International Conference on Robots and Vision (CRV'24), Guelph, Ontario, Canada

Precise Human Mesh Recovery (HMR) with in-the-wild data is a formidable challenge and is often hindered by depth ambiguities and reduced precision. Existing works resort to either pose priors or multi-modal data such as multi-view or point cloud information, though their methods often overlook the valuable scene-depth information inherently present in a single image. Moreover, achieving robust HMR for out-of-distribution (OOD) data is exceedingly challenging due to inherent variations in pose, shape and depth. Consequently, understanding the underlying distribution becomes a vital subproblem in modeling human forms. Motivated by the need for unambiguous and robust human modeling, we introduce Distribution and depth-aware human mesh recovery (D2A-HMR), an end-to-end transformer architecture meticulously designed to minimize the disparity between distributions and incorporate scene-depth leveraging prior depth information. Our approach demonstrates superior performance in handling OOD data in certain scenarios while consistently achieving competitive results against state-of-the-art HMR methods on controlled datasets., Comment: Submitted to 21st International Conference on Robots and Vision (CRV'24), Guelph, Ontario, Canada

Motion segmentation is a fundamental problem in computer vision and is crucial in various applications such as robotics, autonomous driving and action recognition. Recently, spectral clustering based methods have shown impressive results on motion segmentation in dynamic environments. These methods perform spectral clustering on motion affinity matrices to cluster objects or point trajectories in the scene into different motion groups. However, existing methods often need the number of motions present in the scene to be known, which significantly reduces their practicality. In this paper, we propose a unified model selection technique to automatically infer the number of motion groups for spectral clustering based motion segmentation methods by combining different existing model selection techniques together. We evaluate our method on the KT3DMoSeg dataset and achieve competitve results comparing to the baseline where the number of clusters is given as ground truth information., Comment: for the published version, see https://openjournals.uwaterloo.ca/index.php/vsl/article/view/5870/5922

Identifying and segmenting moving objects from a moving monocular camera is difficult when there is unknown camera motion, different types of object motions and complex scene structures. To tackle these challenges, we take advantage of two popular branches of monocular motion segmentation approaches: point trajectory based and optical flow based methods, by synergistically fusing these two highly complementary motion cues at object level. By doing this, we are able to model various complex object motions in different scene structures at once, which has not been achieved by existing methods. We first obtain object-specific point trajectories and optical flow mask for each common object in the video, by leveraging the recent foundational models in object recognition, segmentation and tracking. We then construct two robust affinity matrices representing the pairwise object motion affinities throughout the whole video using epipolar geometry and the motion information provided by optical flow. Finally, co-regularized multi-view spectral clustering is used to fuse the two affinity matrices and obtain the final clustering. Our method shows state-of-the-art performance on the KT3DMoSeg dataset, which contains complex motions and scene structures. Being able to identify moving objects allows us to remove them for map building when using visual SLAM or SFM., Comment: Accepted by IROS 2023 Workshop on Robotic Perception And Mapping: Frontier Vision and Learning Techniques

Player identification is a crucial component in vision-driven soccer analytics, enabling various downstream tasks such as player assessment, in-game analysis, and broadcast production. However, automatically detecting jersey numbers from player tracklets in videos presents challenges due to motion blur, low resolution, distortions, and occlusions. Existing methods, utilizing Spatial Transformer Networks, CNNs, and Vision Transformers, have shown success in image data but struggle with real-world video data, where jersey numbers are not visible in most of the frames. Hence, identifying frames that contain the jersey number is a key sub-problem to tackle. To address these issues, we propose a robust keyframe identification module that extracts frames containing essential high-level information about the jersey number. A spatio-temporal network is then employed to model spatial and temporal context and predict the probabilities of jersey numbers in the video. Additionally, we adopt a multi-task loss function to predict the probability distribution of each digit separately. Extensive evaluations on the SoccerNet dataset demonstrate that incorporating our proposed keyframe identification module results in a significant 37.81% and 37.70% increase in the accuracies of 2 different test sets with domain gaps. These results highlight the effectiveness and importance of our approach in tackling the challenges of automatic jersey number detection in sports videos., Comment: Accepted in the 6th International Workshop on Multimedia Content Analysis in Sports (MMSports'23) @ ACM Multimedia

Using videos to analyze pitchers in baseball can play a vital role in strategizing and injury prevention. Computer vision-based pose analysis offers a time-efficient and cost-effective approach. However, the use of accessible broadcast videos, with a 30fps framerate, often results in partial body motion blur during fast actions, limiting the performance of existing pose keypoint estimation models. Previous works have primarily relied on fixed backgrounds, assuming minimal motion differences between frames, or utilized multiview data to address this problem. To this end, we propose a synthetic data augmentation pipeline to enhance the model's capability to deal with the pitcher's blurry actions. In addition, we leverage in-the-wild videos to make our model robust under different real-world conditions and camera positions. By carefully optimizing the augmentation parameters, we observed a notable reduction in the loss by 54.2% and 36.2% on the test dataset for 2D and 3D pose estimation respectively. By applying our approach to existing state-of-the-art pose estimators, we demonstrate an average improvement of 29.2%. The findings highlight the effectiveness of our method in mitigating the challenges posed by motion blur, thereby enhancing the overall quality of pose estimation., Comment: Accepted in the 6th International Workshop on Multimedia Content Analysis in Sports (MMSports'23) @ ACM Multimedia

The recent success of hybrid methods in monocular odometry has led to many attempts to generalize the performance gains to hybrid monocular SLAM. However, most attempts fall short in several respects, with the most prominent issue being the need for two different map representations (local and global maps), with each requiring different, computationally expensive, and often redundant processes to maintain. Moreover, these maps tend to drift with respect to each other, resulting in contradicting pose and scene estimates, and leading to catastrophic failure. In this paper, we propose a novel approach that makes use of descriptor sharing to generate a single inverse depth scene representation. This representation can be used locally, queried globally to perform loop closure, and has the ability to re-activate previously observed map points after redundant points are marginalized from the local map, eliminating the need for separate and redundant map maintenance processes. The maps generated by our method exhibit no drift between each other, and can be computed at a fraction of the computational cost and memory footprint required by other monocular SLAM systems. Despite the reduced resource requirements, the proposed approach maintains its robustness and accuracy, delivering performance comparable to state-of-the-art SLAM methods (e.g., LDSO, ORB-SLAM3) on the majority of sequences from well-known datasets like EuRoC, KITTI, and TUM VI. The source code is available at: https://github.com/AUBVRL/fslam_ros_docker.

Tracking and identifying players is an important problem in computer vision based ice hockey analytics. Player tracking is a challenging problem since the motion of players in hockey is fast-paced and non-linear. There is also significant player-player and player-board occlusion, camera panning and zooming in hockey broadcast video. Prior published research perform player tracking with the help of handcrafted features for player detection and re-identification. Although commercial solutions for hockey player tracking exist, to the best of our knowledge, no network architectures used, training data or performance metrics are publicly reported. There is currently no published work for hockey player tracking making use of the recent advancements in deep learning while also reporting the current accuracy metrics used in literature. Therefore, in this paper, we compare and contrast several state-of-the-art tracking algorithms and analyze their performance and failure modes in ice hockey., Comment: Accepted to Link\"oping Hockey Analytics Conference (LINHAC). arXiv admin note: substantial text overlap with arXiv:2110.03090

Signage is everywhere and a robot should be able to take advantage of signs to help it localize (including Visual Place Recognition (VPR)) and map. Robust text detection & recognition in the wild is challenging due to such factors as pose, irregular text, illumination, and occlusion. We propose an end-to-end scene text spotting model that simultaneously outputs the text string and bounding boxes. This model is more suitable for VPR. Our central contribution is introducing utilizing an end-to-end scene text spotting framework to adequately capture the irregular and occluded text regions in different challenging places. To evaluate our proposed architecture's performance for VPR, we conducted several experiments on the challenging Self-Collected Text Place (SCTP) benchmark dataset. The initial experimental results show that the proposed method outperforms the SOTA methods in terms of precision and recall when tested on this benchmark., Comment: 6 papged, VI section, typos corrected, revison changes, no result changes

Recent text detection frameworks require several handcrafted components such as anchor generation, non-maximum suppression (NMS), or multiple processing stages (e.g. label generation) to detect arbitrarily shaped text images. In contrast, we propose an end-to-end trainable architecture based on Detection using Transformers (DETR), that outperforms previous state-of-the-art methods in arbitrary-shaped text detection. At its core, our proposed method leverages a bounding box loss function that accurately measures the arbitrary detected text regions' changes in scale and aspect ratio. This is possible due to a hybrid shape representation made from Bezier curves, that are further split into piece-wise polygons. The proposed loss function is then a combination of a generalized-split-intersection-over-union loss defined over the piece-wise polygons and regularized by a Smooth-$\ln$ regression over the Bezier curve's control points. We evaluate our proposed model using Total-Text and CTW-1500 datasets for curved text, and MSRA-TD500 and ICDAR15 datasets for multi-oriented text, and show that the proposed method outperforms the previous state-of-the-art methods in arbitrary-shape text detection tasks.

Causal discovery between collections of time-series data can help diagnose causes of symptoms and hopefully prevent faults before they occur. However, reliable causal discovery can be very challenging, especially when the data acquisition rate varies (i.e., non-uniform data sampling), or in the presence of missing data points (e.g., sparse data sampling). To address these issues, we proposed a new system comprised of two parts, the first part fills missing data with a Gaussian Process Regression, and the second part leverages an Echo State Network, which is a type of reservoir computer (i.e., used for chaotic system modelling) for Causal discovery. We evaluate the performance of our proposed system against three other off-the-shelf causal discovery algorithms, namely, structural expectation-maximization, sub-sampled linear auto-regression absolute coefficients, and multivariate Granger Causality with vector auto-regressive using the Tennessee Eastman chemical dataset; we report on their corresponding Matthews Correlation Coefficient(MCC) and Receiver Operating Characteristic curves (ROC) and show that the proposed system outperforms existing algorithms, demonstrating the viability of our approach to discover causal relationships in a complex system with missing entries.

Identifying players in video is a foundational step in computer vision-based sports analytics. Obtaining player identities is essential for analyzing the game and is used in downstream tasks such as game event recognition. Transformers are the existing standard in Natural Language Processing (NLP) and are swiftly gaining traction in computer vision. Motivated by the increasing success of transformers in computer vision, in this paper, we introduce a transformer network for recognizing players through their jersey numbers in broadcast National Hockey League (NHL) videos. The transformer takes temporal sequences of player frames (also called player tracklets) as input and outputs the probabilities of jersey numbers present in the frames. The proposed network performs better than the previous benchmark on the dataset used. We implement a weakly-supervised training approach by generating approximate frame-level labels for jersey number presence and use the frame-level labels for faster training. We also utilize player shifts available in the NHL play-by-play data by reading the game time using optical character recognition (OCR) to get the players on the ice rink at a certain game time. Using player shifts improved the player identification accuracy by 6%., Comment: CVSports 2022

Tracking and identifying players is a fundamental step in computer vision-based ice hockey analytics. The data generated by tracking is used in many other downstream tasks, such as game event detection and game strategy analysis. Player tracking and identification is a challenging problem since the motion of players in hockey is fast-paced and non-linear when compared to pedestrians. There is also significant camera panning and zooming in hockey broadcast video. Identifying players in ice hockey is challenging since the players of the same team look almost identical, with the jersey number the only discriminating factor between players. In this paper, an automated system to track and identify players in broadcast NHL hockey videos is introduced. The system is composed of three components (1) Player tracking, (2) Team identification and (3) Player identification. Due to the absence of publicly available datasets, the datasets used to train the three components are annotated manually. Player tracking is performed with the help of a state of the art tracking algorithm obtaining a Multi-Object Tracking Accuracy (MOTA) score of 94.5%. For team identification, the away-team jerseys are grouped into a single class and home-team jerseys are grouped in classes according to their jersey color. A convolutional neural network is then trained on the team identification dataset. The team identification network gets an accuracy of 97% on the test set. A novel player identification model is introduced that utilizes a temporal one-dimensional convolutional network to identify players from player bounding box sequences. The player identification model further takes advantage of the available NHL game roster data to obtain a player identification accuracy of 83%., Comment: Submitted to Expert Systems with Applications

Identifying players in sports videos by recognizing their jersey numbers is a challenging task in computer vision. We have designed and implemented a multi-task learning network for jersey number recognition. In order to train a network to recognize jersey numbers, two output label representations are used (1) Holistic - considers the entire jersey number as one class, and (2) Digit-wise - considers the two digits in a jersey number as two separate classes. The proposed network learns both holistic and digit-wise representations through a multi-task loss function. We determine the optimal weights to be assigned to holistic and digit-wise losses through an ablation study. Experimental results demonstrate that the proposed multi-task learning network performs better than the constituent holistic and digit-wise single-task learning networks., Comment: Accepted to the 4th International ACM Workshop on Multimedia Content Analysis in Sports

Puck localization is an important problem in ice hockey video analytics useful for analyzing the game, determining play location, and assessing puck possession. The problem is challenging due to the small size of the puck, excessive motion blur due to high puck velocity and occlusions due to players and boards. In this paper, we introduce and implement a network for puck localization in broadcast hockey video. The network leverages expert NHL play-by-play annotations and uses temporal context to locate the puck. Player locations are incorporated into the network through an attention mechanism by encoding player positions with a Gaussian-based spatial heatmap drawn at player positions. Since event occurrence on the rink and puck location are related, we also perform event recognition by augmenting the puck localization network with an event recognition head and training the network through multi-task learning. Experimental results demonstrate that the network is able to localize the puck with an AUC of $73.1 \%$ on the test set. The puck location can be inferred in 720p broadcast videos at $5$ frames per second. It is also demonstrated that multi-task learning with puck location improves event recognition accuracy., Comment: Accepted at CVSports 2021

In this work, an automatic and simple framework for hockey ice-rink localization from broadcast videos is introduced. First, video is broken into video-shots by a hierarchical partitioning of the video frames, and thresholding based on their histograms. To localize the frames on the ice-rink model, a ResNet18-based regressor is implemented and trained, which regresses to four control points on the model in a frame-by-frame fashion. This leads to the projection jittering problem in the video. To overcome this, in the inference phase, the trajectory of the control points on the ice-rink model are smoothed, for all the consecutive frames of a given video-shot, by convolving a Hann window with the achieved coordinates. Finally, the smoothed homography matrix is computed by using the direct linear transform on the four pairs of corresponding points. A hockey dataset for training and testing the regressor is gathered. The results show success of this simple and comprehensive procedure for localizing the hockey ice-rink and addressing the problem of jittering without affecting the accuracy of homography estimation., Comment: 6 pages, 2 figures, Poster in WiCV2021

Detection and recognition of text in natural images are two main problems in the field of computer vision that have a wide variety of applications in analysis of sports videos, autonomous driving, industrial automation, to name a few. They face common challenging problems that are factors in how text is represented and affected by several environmental conditions. The current state-of-the-art scene text detection and/or recognition methods have exploited the witnessed advancement in deep learning architectures and reported a superior accuracy on benchmark datasets when tackling multi-resolution and multi-oriented text. However, there are still several remaining challenges affecting text in the wild images that cause existing methods to underperform due to there models are not able to generalize to unseen data and the insufficient labeled data. Thus, unlike previous surveys in this field, the objectives of this survey are as follows: first, offering the reader not only a review on the recent advancement in scene text detection and recognition, but also presenting the results of conducting extensive experiments using a unified evaluation framework that assesses pre-trained models of the selected methods on challenging cases, and applies the same evaluation criteria on these techniques. Second, identifying several existing challenges for detecting or recognizing text in the wild images, namely, in-plane-rotation, multi-oriented and multi-resolution text, perspective distortion, illumination reflection, partial occlusion, complex fonts, and special characters. Finally, the paper also presents insight into the potential research directions in this field to address some of the mentioned challenges that are still encountering scene text detection and recognition techniques.

In problems such as sports video analytics, it is difficult to obtain accurate frame level annotations and exact event duration because of the lengthy videos and sheer volume of video data. This issue is even more pronounced in fast-paced sports such as ice hockey. Obtaining annotations on a coarse scale can be much more practical and time efficient. We propose the task of event detection in coarsely annotated videos. We introduce a multi-tower temporal convolutional network architecture for the proposed task. The network, with the help of multiple receptive fields, processes information at various temporal scales to account for the uncertainty with regard to the exact event location and duration. We demonstrate the effectiveness of the multi-receptive field architecture through appropriate ablation studies. The method is evaluated on two tasks - event detection in coarsely annotated hockey videos in the NHL dataset and event spotting in soccer on the SoccerNet dataset. The two datasets lack frame-level annotations and have very distinct event frequencies. Experimental results demonstrate the effectiveness of the network by obtaining a 55% average F1 score on the NHL dataset and by achieving competitive performance compared to the state of the art on the SoccerNet dataset. We believe our approach will help develop more practical pipelines for event detection in sports video.

Anomaly detection refers to the task of finding unusual instances that stand out from the normal data. In several applications, these outliers or anomalous instances are of greater interest compared to the normal ones. Specifically in the case of industrial optical inspection and infrastructure asset management, finding these defects (anomalous regions) is of extreme importance. Traditionally and even today this process has been carried out manually. Humans rely on the saliency of the defects in comparison to the normal texture to detect the defects. However, manual inspection is slow, tedious, subjective and susceptible to human biases. Therefore, the automation of defect detection is desirable. But for defect detection lack of availability of a large number of anomalous instances and labelled data is a problem. In this paper, we present a convolutional auto-encoder architecture for anomaly detection that is trained only on the defect-free (normal) instances. For the test images, residual masks that are obtained by subtracting the original image from the auto-encoder output are thresholded to obtain the defect segmentation masks. The approach was tested on two data-sets and achieved an impressive average F1 score of 0.885. The network learnt to detect the actual shape of the defects even though no defected images were used during the training.

Puck location in ice hockey is essential for hockey analysts for determining the location of play and analyzing game events. However, because of the difficulty involved in obtaining accurate annotations due to the extremely low visibility and commonly occurring occlusions of the puck, the problem is very challenging. The problem becomes even more challenging in broadcast videos with changing camera angles. We introduce a novel methodology for determining puck location from approximate puck location annotations in broadcast video. Our method uniquely leverages the existing puck location information that is publicly available in existing hockey event data and uses the corresponding one-second broadcast video clips as input to the network. The rationale behind using video as input instead of static images is that with video, the temporal information can be utilized to handle puck occlusions. The network outputs a heatmap representing the probability of the puck location using a 3D CNN based architecture. The network is able to regress the puck location from broadcast hockey video clips with varying camera angles. Experimental results demonstrate the capability of the method, achieving 47.07% AUC on the test dataset. The network is also able to estimate the puck location in defensive/offensive zones with an accuracy of greater than 80%.

Humans can easily detect a defect (anomaly) because it is different or salient when compared to the surface it resides on. Today, manual human visual inspection is still the norm because it is difficult to automate anomaly detection. Neural networks are a useful tool that can teach a machine to find defects. However, they require a lot of training examples to learn what a defect is and it is tedious and expensive to get these samples. We tackle the problem of teaching a network with a low number of training samples with a system we call AnoNet. AnoNet's architecture is similar to CompactCNN with the exceptions that (1) it is a fully convolutional network and does not use strided convolution; (2) it is shallow and compact which minimizes over-fitting by design; (3) the compact design constrains the size of intermediate features which allows training to be done without image downsizing; (4) the model footprint is low making it suitable for edge computation; and (5) the anomaly can be detected and localized despite the weak labelling. AnoNet learns to detect the underlying shape of the anomalies despite the weak annotation as well as preserves the spatial localization of the anomaly. Pre-seeding AnoNet with an engineered filter bank initialization technique reduces the total samples required for training and also achieves state-of-the-art performance. Compared to the CompactCNN, AnoNet achieved a massive 94% reduction of network parameters from 1.13 million to 64 thousand parameters. Experiments were conducted on four data-sets and results were compared against CompactCNN and DeepLabv3. AnoNet improved the performance on an average across all data-sets by 106% to an F1 score of 0.98 and by 13% to an AUROC value of 0.942. AnoNet can learn from a limited number of images. For one of the data-sets, AnoNet learnt to detect anomalies after a single pass through just 53 training images.

Visual defect assessment is a form of anomaly detection. This is very relevant in finding faults such as cracks and markings in various surface inspection tasks like pavement and automotive parts. The task involves detection of deviation/divergence of anomalous samples from the normal ones. Two of the major challenges in supervised anomaly detection are the lack of labelled training data and the low availability of anomaly instances. Semi-supervised methods which learn the underlying distribution of the normal samples and then measure the deviation/divergence from the estimated model as the anomaly score have limitations in their overall ability to detect anomalies. This paper proposes the application of network-based deep transfer learning using convolutional neural networks (CNNs) for the task of anomaly detection. Single class SVMs have been used in the past with some success, however we hypothesize that deeper networks for single class classification should perform better. Results obtained on established anomaly detection benchmarks as well as on a real-world dataset, show that the proposed method clearly outperforms the existing state-of-the-art methods, by achieving a staggering average area under the receiver operating characteristic curve value of 0.99 for the tested data-sets which is an average improvement of 41% on the CIFAR10, 20% on MNIST and 16% on Cement Crack data-sets.

In this paper, we present a novel approach called KPTransfer for improving modeling performance for keypoint detection deep neural networks via domain transfer between different keypoint subsets. This approach is motivated by the notion that rich contextual knowledge can be transferred between different keypoint subsets representing separate domains. In particular, the proposed method takes into account various keypoint subsets/domains by sequentially adding and removing keypoints. Contextual knowledge is transferred between two separate domains via domain transfer. Experiments to demonstrate the efficacy of the proposed KPTransfer approach were performed for the task of human pose estimation on the MPII dataset, with comparisons against random initialization and frozen weight extraction configurations. Experimental results demonstrate the efficacy of performing domain transfer between two different joint subsets resulting in a PCKh improvement of up to 1.1 over random initialization on joints such as wrists and knee in certain joint splits with an overall PCKh improvement of 0.5. Domain transfer from a different set of joints not only results in improved accuracy but also results in faster convergence because of mutual co-adaptations of weights resulting from the contextual knowledge of the pose from a different set of joints.

Monocular Odometry systems can be broadly categorized as being either Direct, Indirect, or a hybrid of both. While Indirect systems process an alternative image representation to compute geometric residuals, Direct methods process the image pixels directly to generate photometric residuals. Both paradigms have distinct but often complementary properties. This paper presents a Unified Formulation for Visual Odometry, referred to as UFVO, with the following key contributions: (1) a tight coupling of photometric (Direct) and geometric (Indirect) measurements using a joint multi-objective optimization, (2) the use of a utility function as a decision maker that incorporates prior knowledge on both paradigms, (3) descriptor sharing, where a feature can have more than one type of descriptor and its different descriptors are used for tracking and mapping, (4) the depth estimation of both corner features and pixel features within the same map using an inverse depth parametrization, and (5) a corner and pixel selection strategy that extracts both types of information, while promoting a uniform distribution over the image domain. Experiments show that our proposed system can handle large inter-frame motions, inherits the sub-pixel accuracy of direct methods, can run efficiently in real-time, can generate an Indirect map representation at a marginal computational cost when compared to traditional Indirect systems, all while outperforming state of the art in Direct, Indirect and hybrid systems.

Recognizing actions in ice hockey using computer vision poses challenges due to bulky equipment and inadequate image quality. A novel two-stream framework has been designed to improve action recognition accuracy for hockey using three main components. First, pose is estimated via the Part Affinity Fields model to extract meaningful cues from the player. Second, optical flow (using LiteFlowNet) is used to extract temporal features. Third, pose and optical flow streams are fused and passed to fully-connected layers to estimate the hockey player's action. A novel publicly available dataset named HARPET (Hockey Action Recognition Pose Estimation, Temporal) was created, composed of sequences of annotated actions and pose of hockey players including their hockey sticks as an extension of human body pose. Three contributions are recognized. (1) The novel two-stream architecture achieves 85% action recognition accuracy, with the inclusion of optical flows increasing accuracy by about 10%. (2) The unique localization of hand-held objects (e.g., hockey sticks) as part of pose increases accuracy by about 13%. (3) For pose estimation, a bigger and more general dataset, MSCOCO, is successfully used for transfer learning to a smaller and more specific dataset, HARPET, achieving a PCKh of 87%.

Visual Odometry (VO) can be categorized as being either direct or feature based. When the system is calibrated photometrically, and images are captured at high rates, direct methods have shown to outperform feature-based ones in terms of accuracy and processing time; they are also more robust to failure in feature-deprived environments. On the downside, Direct methods rely on heuristic motion models to seed the estimation of camera motion between frames; in the event that these models are violated (e.g., erratic motion), Direct methods easily fail. This paper proposes a novel system entitled FDMO (Feature assisted Direct Monocular Odometry), which complements the advantages of both direct and featured based techniques. FDMO bootstraps indirect feature tracking upon the sub-pixel accurate localized direct keyframes only when failure modes (e.g., large baselines) of direct tracking occur. Control returns back to direct odometry when these conditions are no longer violated. Efficiencies are introduced to help FDMO perform in real time. FDMO shows significant drift (alignment, rotation & scale) reduction when compared to DSO & ORB SLAM when evaluated using the TumMono and EuroC datasets.

Analysis of retinal fundus images is essential for eye-care physicians in the diagnosis, care and treatment of patients. Accurate fundus and/or retinal vessel maps give rise to longitudinal studies able to utilize multimedia image registration and disease/condition status measurements, as well as applications in surgery preparation and biometrics. The segmentation of retinal morphology has numerous applications in assessing ophthalmologic and cardiovascular disease pathologies. The early detection of many such conditions is often the most effective method for reducing patient risk. Computer aided segmentation of the vasculature has proven to be a challenge, mainly due to inconsistencies such as noise and variations in hue and brightness that can greatly reduce the quality of fundus images. This paper presents PixelBNN, a highly efficient deep method for automating the segmentation of fundus morphologies. The model was trained, tested and cross tested on the DRIVE, STARE and CHASE\_DB1 retinal vessel segmentation datasets. Performance was evaluated using G-mean, Mathews Correlation Coefficient and F1-score. The network was 8.5 times faster than the current state-of-the-art at test time and performed comparatively well, considering a 5 to 19 times reduction in information from resizing images during preprocessing., Comment: Manuscript accepted into SPIE Journal of Medical Imaging special section on Radiomics and Deep Learning