Technical Report / Hochschule Bonn-Rhein-Sieg University of Applied Sciences. Department of Computer Science
Publisher: Dean Prof. Dr. Sascha Alda
Hochschule Bonn-Rhein-Sieg University of Applied Sciences, Department of Computer Science
Sankt Augustin, Germany
ISSN 1869-5272
Hochschule Bonn-Rhein-Sieg University of Applied Sciences, Department of Computer Science
Sankt Augustin, Germany
ISSN 1869-5272
Refine
H-BRS Bibliography
- yes (50)
Departments, institutes and facilities
Document Type
- Report (42)
- Master's Thesis (8)
Year of publication
Has Fulltext
- yes (50)
Keywords
- Robotik (6)
- Cutting sticks-Problem (3)
- Teilsummenaufteilung (3)
- Virtuelle Realität (3)
- 3D-Scanner (2)
- Active Learning (2)
- Computer Vision (2)
- Deep Learning (2)
- Forschungsbericht (2)
- Gravitation (2)
02-2023
Neuromorphic computing aims to mimic the computational principles of the brain in silico and has motivated research into event-based vision and spiking neural networks (SNNs). Event cameras (ECs) capture local, independent changes in brightness, and offer superior power consumption, response latencies, and dynamic ranges compared to frame-based cameras. SNNs replicate neuronal dynamics observed in biological neurons and propagate information in sparse sequences of ”spikes”. Apart from biological fidelity, SNNs have demonstrated potential as an alternative to conventional artificial neural networks (ANNs), such as in reducing energy expenditure and inference time in visual classification. Although potentially beneficial for robotics, the novel event-driven and spike-based paradigms remain scarcely explored outside the domain of aerial robots.
To investigate the utility of brain-inspired sensing and data processing in a robotics application, we developed a neuromorphic approach to real-time, online obstacle avoidance on a manipulator with an onboard camera. Our approach adapts high-level trajectory plans with reactive maneuvers by processing emulated event data in a convolutional SNN, decoding neural activations into avoidance motions, and adjusting plans in a dynamic motion primitive formulation. We conducted simulated and real experiments with a Kinova Gen3 arm performing simple reaching tasks involving static and dynamic obstacles. Our implementation was systematically tuned, validated, and tested in sets of distinct task scenarios, and compared to a non-adaptive baseline through formalized quantitative metrics and qualitative criteria.
The neuromorphic implementation facilitated reliable avoidance of imminent collisions in most scenarios, with 84% and 92% median success rates in simulated and real experiments, where the baseline consistently failed. Adapted trajectories were qualitatively similar to baseline trajectories, indicating low impacts on safety, predictability and smoothness criteria. Among notable properties of the SNN were the correlation of processing time with the magnitude of perceived motions (captured in events) and robustness to different event emulation methods. Preliminary tests with a DAVIS346 EC showed similar performance, validating our experimental event emulation method. These results motivate future efforts to incorporate SNN learning, utilize neuromorphic processors, and target other robot tasks to further explore this approach.
01-2023
This thesis investigates the benefit of rubrics for grading short answers using an active learning mechanism. Automating short answer grading using Natural Language Processing (NLP) is one of the active research areas in the education domain. This could save time for the evaluator and invest more time in preparing for the lecture. Most of the research on short answer grading was treated as a similarity task between reference and student answers. However, grading based on reference answers does not account for partial grades and does not provide feedback. Also, the grading is automatic that tries to replace the evaluator. Hence, using rubrics for short answer grading with active learning eliminates the drawbacks mentioned earlier.
Initially, the proposed approach is evaluated on the Mohler dataset, popularly used to benchmark the methodology. This phase is used to determine the parameters for the proposed approach. Therefore, the approach with the selected parameter exceeds the performance of current State-Of-The-Art (SOTA) methods resulting in the Pearson correlation value of 0.63 and Root Mean Square Error (RMSE) of 0.85. The proposed approach has surpassed the SOTA methods by almost 4%.
Finally, the benchmarked approach is used to grade the short answer based on rubrics instead of reference answers. The proposed approach evaluates short answers from Autonomous Mobile Robot (AMR) dataset to provide scores and feedback (formative assessment) based on the rubrics. The average performance of the dataset results in the Pearson correlation value of 0.61 and RMSE of 0.83. Thus, this research has proven that rubrics-based grading achieves formative assessment without compromising performance. In addition, the rubrics have the advantage of generalizability to all answers.
04-2022
In the field of automatic music generation, one of the greatest challenges is the consistent generation of pieces continuously perceived positively by the majority of the audience since there is no objective method to determine the quality of a musical composition. However, composing principles, which have been refined for millennia, have shaped the core characteristics of today's music. A hybrid music generation system, mlmusic, that incorporates various static, music-theory-based methods, as well as data-driven, subsystems, is implemented to automatically generate pieces considered acceptable by the average listener. Initially, a MIDI dataset, consisting of over 100 hand-picked pieces of various styles and complexities, is analysed using basic music theory principles, and the abstracted information is fed into explicitly constrained LSTM networks. For chord progressions, each individual network is specifically trained on a given sequence length, while phrases are created by consecutively predicting the notes' offset, pitch and duration. Using these outputs as a composition's foundation, additional musical elements, along with constrained recurrent rhythmic and tonal patterns, are statically generated. Although no survey regarding the pieces' reception could be carried out, the successful generation of numerous compositions of varying complexities suggests that the integration of these fundamentally distinctive approaches might lead to success in other branches.
03-2022
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.
02-2022
Recent advances in Natural Language Processing have substantially improved contextualized representations of language. However, the inclusion of factual knowledge, particularly in the biomedical domain, remains challenging. Hence, many Language Models (LMs) are extended by Knowledge Graphs (KGs), but most approaches require entity linking (i.e., explicit alignment between text and KG entities). Inspired by single-stream multimodal Transformers operating on text, image and video data, this thesis proposes the Sophisticated Transformer trained on biomedical text and Knowledge Graphs (STonKGs). STonKGs incorporates a novel multimodal architecture based on a cross encoder that uses the attention mechanism on a concatenation of input sequences derived from text and KG triples, respectively. Over 13 million so-called text-triple pairs, coming from PubMed and assembled using the Integrated Network and Dynamical Reasoning Assembler (INDRA), were used in an unsupervised pre-training procedure to learn representations of biomedical knowledge in STonKGs. By comparing STonKGs to an NLP- and a KG-baseline (operating on either text or KG data) on a benchmark consisting of eight fine-tuning tasks, the proposed knowledge integration method applied in STonKGs was empirically validated. Specifically, on tasks with a comparatively small dataset size and a larger number of classes, STonKGs resulted in considerable performance gains, beating the F1-score of the best baseline by up to 0.083. Both the source code as well as the code used to implement STonKGs are made publicly available so that the proposed method of this thesis can be extended to many other biomedical applications.
01-2022
Effective Neighborhood Feature Exploitation in Graph CNNs for Point Cloud Object-Part Segmentation
(2022)
Part segmentation is the task of semantic segmentation applied on objects and carries a wide range of applications from robotic manipulation to medical imaging. This work deals with the problem of part segmentation on raw, unordered point clouds of 3D objects. While pioneering works on deep learning for point clouds typically ignore taking advantage of local geometric structure around individual points, the subsequent methods proposed to extract features by exploiting local geometry have not yielded significant improvements either. In order to investigate further, a graph convolutional network (GCN) is used in this work in an attempt to increase the effectiveness of such neighborhood feature exploitation approaches. Most of the previous works also focus only on segmenting complete point cloud data. Considering the impracticality of such approaches, taking into consideration the real world scenarios where complete point clouds are scarcely available, this work proposes approaches to deal with partial point cloud segmentation.
In the attempt to better capture neighborhood features, this work proposes a novel method to learn regional part descriptors which guide and refine the segmentation predictions. The proposed approach helps the network achieve state-of-the-art performance of 86.4% mIoU on the ShapeNetPart dataset for methods which do not use any preprocessing techniques or voting strategies. In order to better deal with partial point clouds, this work also proposes new strategies to train and test on partial data. While achieving significant improvements compared to the baseline performance, the problem of partial point cloud segmentation is also viewed through an alternate lens of semantic shape completion.
Semantic shape completion networks not only help deal with partial point cloud segmentation but also enrich the information captured by the system by predicting complete point clouds with corresponding semantic labels for each point. To this end, a new network architecture for semantic shape completion is also proposed based on point completion network (PCN) which takes advantage of a graph convolution based hierarchical decoder for completion as well as segmentation. In addition to predicting complete point clouds, results indicate that the network is capable of reaching within a margin of 5% to the mIoU performance of dedicated segmentation networks for partial point cloud segmentation.
05-2020
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.
04-2020
A Comparative Study of Uncertainty Estimation Methods in Deep Learning Based Classification Models
(2020)
Deep learning models produce overconfident predictions even for misclassified data. This work aims to improve the safety guarantees of software-intensive systems that use deep learning based classification models for decision making by performing comparative evaluation of different uncertainty estimation methods to identify possible misclassifications.
In this work, uncertainty estimation methods applicable to deep learning models are reviewed and those which can be seamlessly integrated to existing deployed deep learning architectures are selected for evaluation. The different uncertainty estimation methods, deep ensembles, test-time data augmentation and Monte Carlo dropout with its variants, are empirically evaluated on two standard datasets (CIFAR-10 and CIFAR-100) and two custom classification datasets (optical inspection and RoboCup@Work dataset). A relative ranking between the methods is provided by evaluating the deep learning classifiers on various aspects such as uncertainty quality, classifier performance and calibration. Standard metrics like entropy, cross-entropy, mutual information, and variance, combined with a rank histogram based method to identify uncertain predictions by thresholding on these metrics, are used to evaluate uncertainty quality.
The results indicate that Monte Carlo dropout combined with test-time data augmentation outperforms all other methods by identifying more than 95% of the misclassifications and representing uncertainty in the highest number of samples in the test set. It also yields a better classifier performance and calibration in terms of higher accuracy and lower Expected Calibration Error (ECE), respectively. A python based uncertainty estimation library for training and real-time uncertainty estimation of deep learning based classification models is also developed.
03-2020
Human and robot tasks in household environments include actions such as carrying an object, cleaning a surface, etc. These tasks are performed by means of dexterous manipulation, and for humans, they are straightforward to accomplish. Moreover, humans perform these actions with reasonable accuracy and precision but with much less energy and stress on the actuators (muscles) than the robots do. The high agility in controlling their forces and motions is actually due to "laziness", i.e. humans exploit the existing natural forces and constraints to execute the tasks.
The above-mentioned properties of the human lazy strategy motivate us to relax the problem of controlling robot motions and forces, and solve it with the help of the environment. Therefore, in this work, we developed a lazy control strategy, i.e. task specification models and control architectures that relax several aspects of robot control by exploiting prior knowledge about the task and environment. The developed control strategy is realized in four different robotics use cases. In this work, the Popov-Vereshchagin hybrid dynamics solver is used as one of the building blocks in the proposed control architectures. An extension of the solver’s interface with the artificial Cartesian force and feed-forward joint torque task-drivers is proposed in this thesis.
To validate the proposed lazy control approach, an experimental evaluation was performed in a simulation environment and on a real robot platform.
02-2020
Object detectors have improved considerably in the last years by using advanced Convolutional Neural Networks (CNNs) architectures. However, many detector hyper-parameters are not generally tuned, and they are used with values set by the detector authors. Blackbox optimization methods have gained more attention in recent years because of its ability to optimize the hyper-parameters of various machine learning algorithms and deep learning models. However, these methods are not explored in improving CNN-based object detector's hyper-parameters. In this research work, we propose the use of blackbox optimization methods such as Gaussian Process based Bayesian Optimization (BOGP), Sequential Model-based Algorithm Configuration (SMAC), and Covariance Matrix Adaptation Evolution Strategy (CMA-ES) to tune the hyper-parameters in Faster R-CNN and Single Shot MultiBox Detector (SSD). In Faster R-CNN, tuning the input image size, prior box anchor scales and ratios using BOGP, SMAC, and CMA-ES has increased the performance around 1.5% in terms of Mean Average Precision (mAP) on PASCAL VOC. Tuning the anchor scales of SSD has increased the mAP by 3% on PASCAL VOC and marine debris datasets. On the COCO dataset with SSD, mAP improvement is observed in the medium and large objects, but mAP decreases by 1% in small objects. The experimental results show that the blackbox optimization methods have proved to increase the mAP performance by optimizing the object detectors. Moreover, it has achieved better results than the hand-tuned configurations in most of the cases.