Refine
H-BRS Bibliography
- yes (30) (remove)
Departments, institutes and facilities
Document Type
- Master's Thesis (30) (remove)
Year of publication
Language
- English (30) (remove)
Keywords
- Active Learning (2)
- Computer Vision (2)
- Emergency support system (2)
- Mobile sensors (2)
- Object Detection (2)
- deep learning (2)
- object detection (2)
- 3D-Scanner (1)
- ASAG (1)
- Adaptive mesh refinement (1)
In order to help journalists investigate inside large audiovisual archives, as maintained by news broadcast agencies, the multimedia data must be indexed by text-based search engies. By automatically creating a transcript through automatic speech recognition (ASR), the spoken word becomes accessible to text search, and queries for keywords are made possible. But stil, important contextual information like the identity of the speaker is not captured. Especially when gathering original footage in the political domain, the identity of the speaker can be the most important query constraint, although this name may not be prominent in the words spoken. It is thus desireable to have this information provided explicitely to the search engine. To provide this information, the archive must be an alyzed by automatic Speaker Identification (SID). While this research topic has seen substantial gains in accuracy and robustness over last years, it has not yet established itself as a helpful, large-scale tool outside the research community. This thesis sets out to establish a workflow to provide automatic speaker identification. Its application is to help journalists searching on speeches given in the German parliament (Bundestag). This is a contribution to the News-Stream 3.0 project, a BMBF funded research project that addresses accessibility of various data sources for journalists.
Interactive Object Detection
(2019)
The success of state-of-the-art object detection methods depend heavily on the availability of a large amount of annotated image data. The raw image data available from various sources are abundant but non-annotated. Annotating image data is often costly, time-consuming or needs expert help. In this work, a new paradigm of learning called Active Learning is explored which uses user interaction to obtain annotations for a subset of the dataset. The goal of active learning is to achieve superior object detection performance with images that are annotated on demand. To realize active learning method, the trade-off between the effort to annotate (annotation cost) unlabeled data and the performance of object detection model is minimised.
Random Forests based method called Hough Forest is chosen as the object detection model and the annotation cost is calculated as the predicted false positive and false negative rate. The framework is successfully evaluated on two Computer Vision benchmark and two Carl Zeiss custom datasets. Also, an evaluation of RGB, HoG and Deep features for the task is presented.
Experimental results show that using Deep features with Hough Forest achieves the maximum performance. By employing Active Learning, it is demonstrated that performance comparable to the fully supervised setting can be achieved by annotating just 2.5% of the images. To this end, an annotation tool is developed for user interaction during Active Learning.
This thesis proposes a multi-label classification approach using the Multimodal Transformer (MulT) [80] to perform multi-modal emotion categorization on a dataset of oral histories archived at the Haus der Geschichte (HdG). Prior uni-modal emotion classification experiments conducted on the novel HdG dataset provided less than satisfactory results. They uncovered issues such as class imbalance, ambiguities in emotion perception between annotators, and lack of representative training data to perform transfer learning [28]. Hence, the objectives of this thesis were to achieve better results by performing a multi-modal fusion and resolving the problems arising from class imbalance and annotator-induced bias in emotion perception. A further objective was to assess the quality of the novel HdG dataset and benchmark the results using SOTA techniques. Through a literature survey on the challenges, models, and datasets related to multi-modal emotion recognition, we created a methodology utilizing the MulT along with a multi-label classification approach. This approach produced a considerable improvement in the overall emotion recognition by obtaining an average AUC of 0.74 and Balanced-accuracy of 0.70 on the HdG dataset, which is comparable to state-of-the-art (SOTA) results on other datasets. In this manner, we were also able to benchmark the novel HdG dataset as well as introduce a novel multi-annotator learning approach to understand each annotator’s relative strengths and weaknesses for emotion perception. Our evaluation results highlight the potential benefits of the novel multi-annotator learning approach in improving overall performance by resolving the problems arising from annotator-induced bias and variation in the perception of emotions. Complementing these results, we performed a further qualitative analysis of the HdG annotations with a psychologist to study the ambiguities found in the annotations. We conclude that the ambiguities in annotations may have resulted from a combination of several socio-psychological factors and systemic issues associated with the process of creating these annotations. As these problems are also present in most multi-modal emotion recognition datasets, we conclude that the domain could benefit from a set of annotation guidelines to create standardized datasets.
Semantic Image Segmentation Combining Visible and Near-Infrared Channels with Depth Information
(2015)
Image understanding is a vital task in computer vision that has many applications in areas such as robotics, surveillance and the automobile industry. An important precondition for image understanding is semantic image segmentation, i.e. the correct labeling of every image pixel with its corresponding object name or class. This thesis proposes a machine learning approach for semantic image segmentation that uses images from a multi-modal camera rig. It demonstrates that semantic segmentation can be improved by combining different image types as inputs to a convolutional neural network (CNN), when compared to a single-image approach. In this work a multi-channel near-infrared (NIR) image, an RGB image and a depth map are used. The detection of people is further improved by using a skin image that indicates the presence of human skin in the scene and is computed based on NIR information. It is also shown that segmentation accuracy can be enhanced by using a class voting method based on a superpixel pre-segmentation. Models are trained for 10-class, 3-class and binary classification tasks using an original dataset. Compared to the NIR-only approach, average class accuracy is increased by 7% for 10-class, and by 22% for 3-class classification, reaching a total of 48% and 70% accuracy, respectively. The binary classification task, which focuses on the detection of people, achieves a classification accuracy of 95% and true positive rate of 66%. The report at hand describes the proposed approach and the encountered challenges and shows that a CNN can successfully learn and combine features from multi-modal image sets and use them to predict scene labeling.
This report presents an approach on a quadrotor dynamics stabilization based on ICP SLAM. Because the quadrotor lacks sensory information to detect its horizontal drift an additional sensor as Hokuyo-UTM has been used to perform on-line ICP-based SLAM. The obtained position estimates were used in control loops to maintain desired position and orientation of the vehicle. Such attitude parameters as height, yaw and position in space were controlled based on the laser data. As a result the quadrotor demonstrated two significant for autonomous navigation capabilities: performance of on-line SLAMon a flying vehicle and maintaining desired position in 3D space. Visual approach on optical flow based on Pyramid Lucas-Kanade algorithm has been touched and tested in different environmental conditions though hasn't been implemented in the control loop. Also the performance of the Hokuyo laser scanner and the related to it ICP SLAM algorithm have been tested in different environmental conditions indoors, outdoors and in presence of smoke. Results are presented and discussed. The requirement of performing on-line SLAM algorithm and to carry quite heavy equipment for it forced to seek a solution to increase the payload of the quadrotor with its computational power. A new hardware and distributed software architectures are therefore presented in the report.
Object detection concerns the classification and localization of objects in an image. To cope with changes in the environment, such as when new classes are added or a new domain is encountered, the detector needs to update itself with the new information while retaining knowledge learned in the past. Previous works have shown that training the detector solely on new data would produce a severe "forgetting" effect, in which the performance on past tasks deteriorates through each new learning phase. However, in many cases, storing and accessing past data is not possible due to privacy concerns or storage constraints. This project aims to investigate promising continual learning strategies for object detection without storing and accessing past training images and labels. We show that by utilizing the pseudo-background trick to deal with missing labels, and knowledge distillation to deal with missing data, the forgetting effect can be significantly reduced in both class-incremental and domain-incremental scenarios. Furthermore, an integration of a small latent replay buffer can result in a positive backward transfer, indicating the enhancement of past knowledge when new knowledge is learned.
Modern engineering relies heavily on utilizing computer technologies. This is especially true for thermoplastic manufacturing, such as blow molding. A crucial milestone for digitalization is the continuous integration of data in unified or interoperable systems. While new simulation technologies are constantly developed, data management standards such as STEP fail at integrating them. On the other hand, industrial standards such as ”VMAP” manage to improve interoperability for Small and Medium-sized Enterprises. However, they do not provide Simulation Process and Data Management (SPDM) technologies. For SPDM integration of VMAP data, Ontology-Based Data Access is used to allow continuing the digital thread in custom semantic-based open-source solutions. An ontology of the database format (VMAP) was generated alongside an expandable knowledge graph of data access methods. A Python-based software architecture was developed, automatically using the semantic representations of database format and data access to query data and metadata within the VMAP file. The result is a software architecture template that can be adapted for other data standards and integrated into semantic data management systems. It allows semantic queries on simulation data down to element-wise resolution without integrating the whole model information. The architecture can instantiate a file in a knowledge graph, query a file’s metadatum and, in case it is not yet available, find a semantically represented process that allows the creation and instantiation of the required metadatum. See Figure 1. The results of this thesis can be expected to form a basis for semantic SPDM tools.
This project focuses on object detection in dense volume data. There are several types of dense volume data, namely Computed Tomography (CT) scan, Positron Emission Tomography (PET), Magnetic Resonance Imaging (MRI). This work focuses on CT scans. CT scans are not limited to the medical domain; they are also used in industries. CT scans are used in airport baggage screening, assembly lines, and the object detection systems in these places should be able to detect objects fast. One of the ways to address the issue of computational complexity and make the object detection systems fast is to use low-resolution images. Low-resolution CT scanning is fast. The entire process of scanning and detection can be made faster by using low-resolution images. Even in the medical domain, to reduce the rad iation dose, the exposure time of the patient should be reduced. The exposure time of patients could be reduced by allowing low-resolution CT scans. Hence it is essential to find out which object detection model has better accuracy as well as speed at low-resolution CT scans. However, the existing approaches did not provide details about how the model would perform when the resolution of CT scans is varied. Hence in this project, the goal is to analyze the impact of varying resolution of CT scans on both the speed and accuracy of the model. Three object detection models, namely RetinaNet, YOLOv3, and YOLOv5, were trained at various resolutions. Among the three models, it was found that YOLOv5 has the best mAP and f1 score at multiple resolutions on the DeepLesion dataset. RetinaNet model h as the least inference time on the DeepLesion dataset. From the experiments, it could be asserted that sacrificing mean average precision (mAP) to improve inference time by reducing resolution is feasible.
The ability to finely segment different instances of various objects in an environment forms a critical tool in the perception tool-box of any autonomous agent. Traditionally instance segmentation is treated as a multi-label pixel-wise classification problem. This formulation has resulted in networks that are capable of producing high-quality instance masks but are extremely slow for real-world usage, especially on platforms with limited computational capabilities. This thesis investigates an alternate regression-based formulation of instance segmentation to achieve a good trade-off between mask precision and run-time. Particularly the instance masks are parameterized and a CNN is trained to regress to these parameters, analogous to bounding box regression performed by an object detection network.
In this investigation, the instance segmentation masks in the Cityscape dataset are approximated using irregular octagons and an existing object detector network (i.e., SqueezeDet) is modified to regresses to the parameters of these octagonal approximations. The resulting network is referred to as SqueezeDetOcta. At the image boundaries, object instances are only partially visible. Due to the convolutional nature of most object detection networks, special handling of the boundary adhering object instances is warranted. However, the current object detection techniques seem to be unaffected by this and handle all the object instances alike. To this end, this work proposes selectively learning only partial, untainted parameters of the bounding box approximation of the boundary adhering object instances. Anchor-based object detection networks like SqueezeDet and YOLOv2 have a discrepancy between the ground-truth encoding/decoding scheme and the coordinate space used for clustering, to generate the prior anchor shapes. To resolve this disagreement, this work proposes clustering in a space defined by two coordinate axes representing the natural log transformations of the width and height of the ground-truth bounding boxes.
When both SqueezeDet and SqueezeDetOcta were trained from scratch, SqueezeDetOcta lagged behind the SqueezeDet network by a massive ≈ 6.19 mAP. Further analysis revealed that the sparsity of the annotated data was the reason for this lackluster performance of the SqueezeDetOcta network. To mitigate this issue transfer-learning was used to fine-tune the SqueezeDetOcta network starting from the trained weights of the SqueezeDet network. When all the layers of the SqueezeDetOcta were fine-tuned, it outperformed the SqueezeDet network paired with logarithmically extracted anchors by ≈ 0.77 mAP. In addition to this, the forward pass latencies of both SqueezeDet and SqueezeDetOcta are close to ≈ 19ms. Boundary adhesion considerations, during training, resulted in an improvement of ≈ 2.62 mAP of the baseline SqueezeDet network. A SqueezeDet network paired with logarithmically extracted anchors improved the performance of the baseline SqueezeDet network by ≈ 1.85 mAP.
In summary, this work demonstrates that if given sufficient fine instance annotated data, an existing object detection network can be modified to predict much finer approximations (i.e., irregular octagons) of the instance annotations, whilst having the same forward pass latency as that of the bounding box predicting network. The results justify the merits of logarithmically extracted anchors to boost the performance of any anchor-based object detection network. The results also showed that the special handling of image boundary adhering object instances produces more performant object detectors.
As cameras are ubiquitous in autonomous systems, object detection is a crucial task. Object detectors are widely used in applications such as autonomous driving, healthcare, and robotics. Given an image, an object detector outputs both the bounding box coordinates as well as classification probabilities for each object detected. The state-of-the-art detectors are treated as black boxes due to their highly non-linear internal computations. Even with unprecedented advancements in detector performance, the inability to explain how their outputs are generated limits their use in safety-critical applications in particular. It is therefore crucial to explain the reason behind each detector decision in order to gain user trust, enhance detector performance, and analyze their failure.
Previous work fails to explain as well as evaluate both bounding box and classification decisions individually for various detectors. Moreover, no tools explain each detector decision, evaluate the explanations, and also identify the reasons for detector failures. This restricts the flexibility to analyze detectors. The main contribution presented here is an open-source Detector Explanation Toolkit (DExT). It is used to explain the detector decisions, evaluate the explanations, and analyze detector errors. The detector decisions are explained visually by highlighting the image pixels that most influence a particular decision. The toolkit implements the proposed approach to generate a holistic explanation for all detector decisions using certain gradient-based explanation methods. To the author’s knowledge, this is the first work to conduct extensive qualitative and novel quantitative evaluations of different explanation methods across various detectors. The qualitative evaluation incorporates a visual analysis of the explanations carried out by the author as well as a human-centric evaluation. The human-centric evaluation includes a user study to understand user trust in the explanations generated across various explanation methods for different detectors. Four multi-object visualization methods are provided to merge the explanations of multiple objects detected in an image as well as the corresponding detector outputs in a single image. Finally, DExT implements the procedure to analyze detector failures using the formulated approach.
The visual analysis illustrates that the ability to explain a model is more dependent on the model itself than the actual ability of the explanation method. In addition, the explanations are affected by the object explained, the decision explained, detector architecture, training data labels, and model parameters. The results of the quantitative evaluation show that the Single Shot MultiBox Detector (SSD) is more faithfully explained compared to other detectors regardless of the explanation methods. In addition, a single explanation method cannot generate more faithful explanations than other methods for both the bounding box and the classification decision across different detectors. Both the quantitative and human-centric evaluations identify that SmoothGrad with Guided Backpropagation (GBP) provides more trustworthy explanations among selected methods across all detectors. Finally, a convex polygon-based multi-object visualization method provides more human-understandable visualization than other methods.
The author expects that DExT will motivate practitioners to evaluate object detectors from the interpretability perspective by explaining both bounding box and classification decisions.