The success of deep learning in recent years has lead to a rising demand for neural network architecture engineering. As a consequence, neural architecture search (NAS), which aims at automatically designing neural network architectures in a data-driven manner rather than manually, has evolved as a popular field of research. With the advent of weight sharing strategies across architectures, NAS has become applicable to a much wider range of problems. In particular, there are now many publications for dense prediction tasks in computer vision that require pixel-level predictions, such as semantic segmentation or object detection. These tasks come with novel challenges, such as higher memory footprints due to high-resolution data, learning multi-scale representations, longer training times, and more complex and larger neural architectures. In this manuscript, we provide an overview of NAS for dense prediction tasks by elaborating on these novel challenges and surveying ways to address them to ease future research and application of existing methods to novel problems.
As robotic systems become more and more capable of assisting humans in their everyday lives, we must consider the opportunities for these artificial agents to make their human collaborators feel unsafe or to treat them unfairly. Robots can exhibit antisocial behavior causing physical harm to people or reproduce unfair behavior replicating and even amplifying historical and societal biases which are detrimental to humans they interact with. In this paper, we discuss these issues considering sociable robotic manipulation and fair robotic decision making. We propose a novel approach to learning fair and sociable behavior, not by reproducing positive behavior, but rather by avoiding negative behavior. In this study, we highlight the importance of incorporating sociability in robot manipulation, as well as the need to consider fairness in human-robot interactions.
Object recognition for the most part has been approached as a one-hot problem that treats classes to be discrete and unrelated. Each image region has to be assigned to one member of a set of objects, including a background class, disregarding any similarities in the object types. In this work, we compare the error statistics of the class embeddings learned from a one-hot approach with semantically structured embeddings from natural language processing or knowledge graphs that are widely applied in open world object detection. Extensive experimental results on multiple knowledge-embeddings as well as distance metrics indicate that knowledge-based class representations result in more semantically grounded misclassifications while performing on par compared to one-hot methods on the challenging COCO and Cityscapes object detection benchmarks. We generalize our findings to multiple object detection architectures by proposing a knowledge-embedded design for keypoint-based and transformer-based object detection architectures.
In this technical report, we describe our EfficientLPT architecture that won the panoptic tracking challenge in the 7th AI Driving Olympics at NeurIPS 2021. Our architecture builds upon the top-down EfficientLPS panoptic segmentation approach. EfficientLPT consists of a shared backbone with a modified EfficientNet-B5 model comprising the proximity convolution module as the encoder followed by the range-aware FPN to aggregate semantically rich range-aware multi-scale features. Subsequently, we employ two task-specific heads, the scale-invariant semantic head and hybrid task cascade with feedback from the semantic head as the instance head. Further, we employ a novel panoptic fusion module to adaptively fuse logits from each of the heads to yield the panoptic tracking output. Our approach exploits three consecutive accumulated scans to predict locally consistent panoptic tracking IDs and also the overlap between the scans to predict globally consistent panoptic tracking IDs for a given sequence. The benchmarking results from the 7th AI Driving Olympics at NeurIPS 2021 show that our model is ranked #1 for the panoptic tracking task on the Panoptic nuScenes dataset.
Audio-visual navigation combines sight and hearing to navigate to a sound-emitting source in an unmapped environment. While recent approaches have demonstrated the benefits of audio input to detect and find the goal, they focus on clean and static sound sources and struggle to generalize to unheard sounds. In this work, we propose the novel dynamic audio-visual navigation benchmark which requires to catch a moving sound source in an environment with noisy and distracting sounds. We introduce a reinforcement learning approach that learns a robust navigation policy for these complex settings. To achieve this, we propose an architecture that fuses audio-visual information in the spatial feature space to learn correlations of geometric information inherent in both local maps and audio signals. We demonstrate that our approach consistently outperforms the current state-of-the-art by a large margin across all tasks of moving sounds, unheard sounds, and noisy environments, on two challenging 3D scanned real-world environments, namely Matterport3D and Replica. The benchmark is available at http://dav-nav.cs.uni-freiburg.de.
A core challenge for an autonomous agent acting in the real world is to adapt its repertoire of skills to cope with its noisy perception and dynamics. To scale learning of skills to long-horizon tasks, robots should be able to learn and later refine their skills in a structured manner through trajectories rather than making instantaneous decisions individually at each time step. To this end, we propose the Soft Actor-Critic Gaussian Mixture Model (SAC-GMM), a novel hybrid approach that learns robot skills through a dynamical system and adapts the learned skills in their own trajectory distribution space through interactions with the environment. Our approach combines classical robotics techniques of learning from demonstration with the deep reinforcement learning framework and exploits their complementary nature. We show that our method utilizes sensors solely available during the execution of preliminarily learned skills to extract relevant features that lead to faster skill refinement. Extensive evaluations in both simulation and real-world environments demonstrate the effectiveness of our method in refining robot skills by leveraging physical interactions, high-dimensional sensory data, and sparse task completion rewards. Videos, code, and pre-trained models are available at \url{http://sac-gmm.cs.uni-freiburg.de}.
Learning to solve complex manipulation tasks from visual observations is a dominant challenge for real-world robot learning. Deep reinforcement learning algorithms have recently demonstrated impressive results, although they still require an impractical amount of time-consuming trial-and-error iterations. In this work, we consider the promising alternative paradigm of interactive learning where a human teacher provides feedback to the policy during execution, as opposed to imitation learning where a pre-collected dataset of perfect demonstrations is used. Our proposed CEILing (Corrective and Evaluative Interactive Learning) framework combines both corrective and evaluative feedback from the teacher to train a stochastic policy in an asynchronous manner, and employs a dedicated mechanism to trade off human corrections with the robot's own experience. We present results obtained with our framework in extensive simulation and real-world experiments that demonstrate that CEILing can effectively solve complex robot manipulation tasks directly from raw images in less than one hour of real-world training.
Scene understanding is a pivotal task for autonomous vehicles to safely navigate in the environment. Recent advances in deep learning enable accurate semantic reconstruction of the surroundings from LiDAR data. However, these models encounter a large domain gap while deploying them on vehicles equipped with different LiDAR setups which drastically decreases their performance. Fine-tuning the model for every new setup is infeasible due to the expensive and cumbersome process of recording and manually labeling new data. Unsupervised Domain Adaptation (UDA) techniques are thus essential to fill this domain gap and retain the performance of models on new sensor setups without the need for additional data labeling. In this paper, we propose AdaptLPS, a novel UDA approach for LiDAR panoptic segmentation that leverages task-specific knowledge and accounts for variation in the number of scan lines, mounting position, intensity distribution, and environmental conditions. We tackle the UDA task by employing two complementary domain adaptation strategies, data-based and model-based. While data-based adaptations reduce the domain gap by processing the raw LiDAR scans to resemble the scans in the target domain, model-based techniques guide the network in extracting features that are representative for both domains. Extensive evaluations on three pairs of real-world autonomous driving datasets demonstrate that AdaptLPS outperforms existing UDA approaches by up to 6.41 pp in terms of the PQ score.
Panoptic scene understanding and tracking of dynamic agents are essential for robots and automated vehicles to navigate in urban environments. As LiDARs provide accurate illumination-independent geometric depictions of the scene, performing these tasks using LiDAR point clouds provides reliable predictions. However, existing datasets lack diversity in the type of urban scenes and have a limited number of dynamic object instances which hinders both learning of these tasks as well as credible benchmarking of the developed methods. In this paper, we introduce the large-scale Panoptic nuScenes benchmark dataset that extends our popular nuScenes dataset with point-wise groundtruth annotations for semantic segmentation, panoptic segmentation, and panoptic tracking tasks. To facilitate comparison, we provide several strong baselines for each of these tasks on our proposed dataset. Moreover, we analyze the drawbacks of the existing metrics for panoptic tracking and propose the novel instance-centric PAT metric that addresses the concerns. We present exhaustive experiments that demonstrate the utility of Panoptic nuScenes compared to existing datasets and make the online evaluation server available at nuScenes.org. We believe that this extension will accelerate the research of novel methods for scene understanding of dynamic urban environments.
Bird's-Eye-View (BEV) maps have emerged as one of the most powerful representations for scene understanding due to their ability to provide rich spatial context while being easy to interpret and process. However, generating BEV maps requires complex multi-stage paradigms that encapsulate a series of distinct tasks such as depth estimation, ground plane estimation, and semantic segmentation. These sub-tasks are often learned in a disjoint manner which prevents the model from holistic reasoning and results in erroneous BEV maps. Moreover, existing algorithms only predict the semantics in the BEV space, which limits their use in applications where the notion of object instances is critical. In this work, we present the first end-to-end learning approach for directly predicting dense panoptic segmentation maps in the BEV, given a single monocular image in the frontal view (FV). Our architecture follows the top-down paradigm and incorporates a novel dense transformer module consisting of two distinct transformers that learn to independently map vertical and flat regions in the input image from the FV to the BEV. Additionally, we derive a mathematical formulation for the sensitivity of the FV-BEV transformation which allows us to intelligently weight pixels in the BEV space to account for the varying descriptiveness across the FV image. Extensive evaluations on the KITTI-360 and nuScenes datasets demonstrate that our approach exceeds the state-of-the-art in the PQ metric by 3.61 pp and 4.93 pp respectively.