Refine
H-BRS Bibliography
- yes (19)
Departments, institutes and facilities
Document Type
- Preprint (14)
- Conference Object (4)
- Article (1)
Language
- English (19)
Keywords
Reinforcement learning (RL) algorithms should learn as much as possible about the environment but not the properties of the physics engines that generate the environment. There are multiple algorithms that solve the task in a physics engine based environment but there is no work done so far to understand if the RL algorithms can generalize across physics engines. In this work, we compare the generalization performance of various deep reinforcement learning algorithms on a variety of control tasks. Our results show that MuJoCo is the best engine to transfer the learning to other engines. On the other hand, none of the algorithms generalize when trained on PyBullet. We also found out that various algorithms have a promising generalizability if the effect of random seeds can be minimized on their performance.
Traffic sign recognition is an important component of many advanced driving assistance systems, and it is required for full autonomous driving. Computational performance is usually the bottleneck in using large scale neural networks for this purpose. SqueezeNet is a good candidate for efficient image classification of traffic signs, but in our experiments it does not reach high accuracy, and we believe this is due to lack of data, requiring data augmentation. Generative adversarial networks can learn the high dimensional distribution of empirical data, allowing the generation of new data points. In this paper we apply pix2pix GANs architecture to generate new traffic sign images and evaluate the use of these images in data augmentation. We were motivated to use pix2pix to translate symbolic sign images to real ones due to the mode collapse in Conditional GANs. Through our experiments we found that data augmentation using GAN can increase classification accuracy for circular traffic signs from 92.1% to 94.0%, and for triangular traffic signs from 93.8% to 95.3%, producing an overall improvement of 2%. However some traditional augmentation techniques can outperform GAN data augmentation, for example contrast variation in circular traffic signs (95.5%) and displacement on triangular traffic signs (96.7 %). Our negative results shows that while GANs can be naively used for data augmentation, they are not always the best choice, depending on the problem and variability in the data.
Self-supervised learning has proved to be a powerful approach to learn image representations without the need of large labeled datasets. For underwater robotics, it is of great interest to design computer vision algorithms to improve perception capabilities such as sonar image classification. Due to the confidential nature of sonar imaging and the difficulty to interpret sonar images, it is challenging to create public large labeled sonar datasets to train supervised learning algorithms. In this work, we investigate the potential of three self-supervised learning methods (RotNet, Denoising Autoencoders, and Jigsaw) to learn high-quality sonar image representation without the need of human labels. We present pre-training and transfer learning results on real-life sonar image datasets. Our results indicate that self-supervised pre-training yields classification performance comparable to supervised pre-training in a few-shot transfer learning setup across all three methods. Code and self-supervised pre-trained models are be available at https://github.com/agrija9/ssl-sonar-images
Facial emotion recognition is the task to classify human emotions in face images. It is a difficult task due to high aleatoric uncertainty and visual ambiguity. A large part of the literature aims to show progress by increasing accuracy on this task, but this ignores the inherent uncertainty and ambiguity in the task. In this paper we show that Bayesian Neural Networks, as approximated using MC-Dropout, MC-DropConnect, or an Ensemble, are able to model the aleatoric uncertainty in facial emotion recognition, and produce output probabilities that are closer to what a human expects. We also show that calibration metrics show strange behaviors for this task, due to the multiple classes that can be considered correct, which motivates future work. We believe our work will motivate other researchers to move away from Classical and into Bayesian Neural Networks.
In this paper we introduce the Perception for Autonomous Systems (PAZ) software library. PAZ is a hierarchical perception library that allow users to manipulate multiple levels of abstraction in accordance to their requirements or skill level. More specifically, PAZ is divided into three hierarchical levels which we refer to as pipelines, processors, and backends. These abstractions allows users to compose functions in a hierarchical modular scheme that can be applied for preprocessing, data-augmentation, prediction and postprocessing of inputs and outputs of machine learning (ML) models. PAZ uses these abstractions to build reusable training and prediction pipelines for multiple robot perception tasks such as: 2D keypoint estimation, 2D object detection, 3D keypoint discovery, 6D pose estimation, emotion classification, face recognition, instance segmentation, and attention mechanisms.
Machine learning and neural networks are now ubiquitous in sonar perception, but it lags behind the computer vision field due to the lack of data and pre-trained models specifically for sonar images. In this paper we present the Marine Debris Turntable dataset and produce pre-trained neural networks trained on this dataset, meant to fill the gap of missing pre-trained models for sonar images. We train Resnet 20, MobileNets, DenseNet121, SqueezeNet, MiniXception, and an Autoencoder, over several input image sizes, from 32 x 32 to 96 x 96, on the Marine Debris turntable dataset. We evaluate these models using transfer learning for low-shot classification in the Marine Debris Watertank and another dataset captured using a Gemini 720i sonar. Our results show that in both datasets the pre-trained models produce good features that allow good classification accuracy with low samples (10-30 samples per class). The Gemini dataset validates that the features transfer to other kinds of sonar sensors. We expect that the community benefits from the public release of our pre-trained models and the turntable dataset.
State-of-the-art object 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. Previous work fails to produce explanations for both bounding box and classification decisions, and generally make individual explanations for various detectors. In this paper, we propose an open-source Detector Explanation Toolkit (DExT) which implements the proposed approach to generate a holistic explanation for all detector decisions using certain gradient-based explanation methods. We suggests various multi-object visualization methods to merge the explanations of multiple objects detected in an image as well as the corresponding detections in a single image. The quantitative evaluation show that the Single Shot MultiBox Detector (SSD) is more faithfully explained compared to other detectors regardless of the explanation methods. Both quantitative and human-centric evaluations identify that SmoothGrad with Guided Backpropagation (GBP) provides more trustworthy explanations among selected methods across all detectors. We expect that DExT will motivate practitioners to evaluate object detectors from the interpretability perspective by explaining both bounding box and classification decisions.
21 pages, with supplementary
Application of underwater robots are on the rise, most of them are dependent on sonar for underwater vision, but the lack of strong perception capabilities limits them in this task. An important issue in sonar perception is matching image patches, which can enable other techniques like localization, change detection, and mapping. There is a rich literature for this problem in color images, but for acoustic images, it is lacking, due to the physics that produce these images. In this paper we improve on our previous results for this problem (Valdenegro-Toro et al, 2017), instead of modeling features manually, a Convolutional Neural Network (CNN) learns a similarity function and predicts if two input sonar images are similar or not. With the objective of improving the sonar image matching problem further, three state of the art CNN architectures are evaluated on the Marine Debris dataset, namely DenseNet, and VGG, with a siamese or two-channel architecture, and contrastive loss. To ensure a fair evaluation of each network, thorough hyper-parameter optimization is executed. We find that the best performing models are DenseNet Two-Channel network with 0.955 AUC, VGG-Siamese with contrastive loss at 0.949 AUC and DenseNet Siamese with 0.921 AUC. By ensembling the top performing DenseNet two-channel and DenseNet-Siamese models overall highest prediction accuracy obtained is 0.978 AUC, showing a large improvement over the 0.91 AUC in the state of the art.
Saliency methods are frequently used to explain Deep Neural Network-based models. Adebayo et al.'s work on evaluating saliency methods for classification models illustrate certain explanation methods fail the model and data randomization tests. However, on extending the tests for various state of the art object detectors we illustrate that the ability to explain a model is more dependent on the model itself than the explanation method. We perform sanity checks for object detection and define new qualitative criteria to evaluate the saliency explanations, both for object classification and bounding box decisions, using Guided Backpropagation, Integrated Gradients, and their Smoothgrad versions, together with Faster R-CNN, SSD, and EfficientDet-D0, trained on COCO. In addition, the sensitivity of the explanation method to model parameters and data labels varies class-wise motivating to perform the sanity checks for each class. We find that EfficientDet-D0 is the most interpretable method independent of the saliency method, which passes the sanity checks with little problems.
Deep learning models are extensively used in various safety critical applications. Hence these models along with being accurate need to be highly reliable. One way of achieving this is by quantifying uncertainty. Bayesian methods for UQ have been extensively studied for Deep Learning models applied on images but have been less explored for 3D modalities such as point clouds often used for Robots and Autonomous Systems. In this work, we evaluate three uncertainty quantification methods namely Deep Ensembles, MC-Dropout and MC-DropConnect on the DarkNet21Seg 3D semantic segmentation model and comprehensively analyze the impact of various parameters such as number of models in ensembles or forward passes, and drop probability values, on task performance and uncertainty estimate quality. We find that Deep Ensembles outperforms other methods in both performance and uncertainty metrics. Deep ensembles outperform other methods by a margin of 2.4% in terms of mIOU, 1.3% in terms of accuracy, while providing reliable uncertainty for decision making.
Emotion and gender recognition from facial features are important properties of human empathy. Robots should also have these capabilities. For this purpose we have designed special convolutional modules that allow a model to recognize emotions and gender with a considerable lower number of parameters, enabling real-time evaluation on a constrained platform. We report accuracies of 96% in the IMDB gender dataset and 66% in the FER-2013 emotion dataset, while requiring a computation time of less than 0.008 seconds on a Core i7 CPU. All our code, demos and pre-trained architectures have been released under an open-source license in our repository at https://github.com/oarriaga/face classification.
Current robot platforms are being employed to collaborate with humans in a wide range of domestic and industrial tasks. These environments require autonomous systems that are able to classify and communicate anomalous situations such as fires, injured persons, car accidents; or generally, any potentially dangerous situation for humans. In this paper we introduce an anomaly detection dataset for the purpose of robot applications as well as the design and implementation of a deep learning architecture that classifies and describes dangerous situations using only a single image as input. We report a classification accuracy of 97 % and METEOR score of 16.2. We will make the dataset publicly available after this paper is accepted.
In this paper we propose an implement a general convolutional neural network (CNN) building framework for designing real-time CNNs. We validate our models by creating a real-time vision system which accomplishes the tasks of face detection, gender classification and emotion classification simultaneously in one blended step using our proposed CNN architecture. After presenting the details of the training procedure setup we proceed to evaluate on standard benchmark sets. We report accuracies of 96% in the IMDB gender dataset and 66% in the FER-2013 emotion dataset. Along with this we also introduced the very recent real-time enabled guided back-propagation visualization technique. Guided back-propagation uncovers the dynamics of the weight changes and evaluates the learned features. We argue that the careful implementation of modern CNN architectures, the use of the current regularization methods and the visualization of previously hidden features are necessary in order to reduce the gap between slow performances and real-time architectures. Our system has been validated by its deployment on a Care-O-bot 3 robot used during RoboCup@Home competitions. All our code, demos and pre-trained architectures have been released under an open-source license in our public repository.
Object detectors have improved considerably in the last years by using advanced CNN architectures. However, many detector hyper-parameters are generally manually tuned, or they are used with values set by the detector authors. Automatic Hyper-parameter optimization has not been explored in improving CNN-based object detectors hyper-parameters. In this work, we propose the use of Black-box optimization methods to tune the prior/default box scales in Faster R-CNN and SSD, using Bayesian Optimization, SMAC, and CMA-ES. We show that by tuning the input image size and prior box anchor scale on Faster R-CNN mAP increases by 2% on PASCAL VOC 2007, and by 3% with SSD. On the COCO dataset with SSD there are mAP improvement in the medium and large objects, but mAP decreases by 1% in small objects. We also perform a regression analysis to find the significant hyper-parameters to tune.
Safety-critical applications like autonomous driving use Deep Neural Networks (DNNs) for object detection and segmentation. The DNNs fail to predict when they observe an Out-of-Distribution (OOD) input leading to catastrophic consequences. Existing OOD detection methods were extensively studied for image inputs but have not been explored much for LiDAR inputs. So in this study, we proposed two datasets for benchmarking OOD detection in 3D semantic segmentation. We used Maximum Softmax Probability and Entropy scores generated using Deep Ensembles and Flipout versions of RandLA-Net as OOD scores. We observed that Deep Ensembles out perform Flipout model in OOD detection with greater AUROC scores for both datasets.