Cornelia Fermüller

Levi Burner, Anton Mitrokhin, Cornelia Fermüller et al.

A new event camera dataset, EVIMO2, is introduced that improves on the popular EVIMO dataset by providing more data, from better cameras, in more complex scenarios. As with its predecessor, EVIMO2 provides labels in the form of per-pixel ground truth depth and segmentation as well as camera and object poses. All sequences use data from physical cameras and many sequences feature multiple independently moving objects. Typically, such labeled data is unavailable in physical event camera datasets. Thus, EVIMO2 will serve as a challenging benchmark for existing algorithms and rich training set for the development of new algorithms. In particular, EVIMO2 is suited for supporting research in motion and object segmentation, optical flow, structure from motion, and visual (inertial) odometry in both monocular or stereo configurations. EVIMO2 consists of 41 minutes of data from three 640$\times$480 event cameras, one 2080$\times$1552 classical color camera, inertial measurements from two six axis inertial measurement units, and millimeter accurate object poses from a Vicon motion capture system. The dataset's 173 sequences are arranged into three categories. 3.75 minutes of independently moving household objects, 22.55 minutes of static scenes, and 14.85 minutes of basic motions in shallow scenes. Some sequences were recorded in low-light conditions where conventional cameras fail. Depth and segmentation are provided at 60 Hz for the event cameras and 30 Hz for the classical camera. The masks can be regenerated using open-source code up to rates as high as 200 Hz. This technical report briefly describes EVIMO2. The full documentation is available online. Videos of individual sequences can be sampled on the download page.

Snehesh Shrestha, Cornelia Fermüller, Tianyu Huang et al.

Pose Estimation techniques rely on visual cues available through observations represented in the form of pixels. But the performance is bounded by the frame rate of the video and struggles from motion blur, occlusions, and temporal coherence. This issue is magnified when people are interacting with objects and instruments, for example playing the violin. Standard approaches for postprocessing use interpolation and smoothing functions to filter noise and fill gaps, but they cannot model highly non-linear motion. We present a method that leverages our understanding of the high degree of a causal relationship between the sound produced and the motion that produces them. We use the audio signature to refine and predict accurate human body pose motion models. We propose MAPnet (Music Assisted Pose network) for generating a fine grain motion model from sparse input pose sequences but continuous audio. To accelerate further research in this domain, we also open-source MAPdat, a new multi-modal dataset of 3D violin playing motion with music. We perform a comparison of different standard machine learning models and perform analysis on input modalities, sampling techniques, and audio and motion features. Experiments on MAPdat suggest multi-modal approaches like ours as a promising direction for tasks previously approached with visual methods only. Our results show both qualitatively and quantitatively how audio can be combined with visual observation to help improve any pose estimation methods.

Chahat Deep Singh, Riya Kumari, Cornelia Fermüller et al.

In the era of deep learning, data is the critical determining factor in the performance of neural network models. Generating large datasets suffers from various difficulties such as scalability, cost efficiency and photorealism. To avoid expensive and strenuous dataset collection and annotations, researchers have inclined towards computer-generated datasets. Although, a lack of photorealism and a limited amount of computer-aided data, has bounded the accuracy of network predictions. To this end, we present WorldGen -- an open source framework to autonomously generate countless structured and unstructured 3D photorealistic scenes such as city view, object collection, and object fragmentation along with its rich ground truth annotation data. WorldGen being a generative model gives the user full access and control to features such as texture, object structure, motion, camera and lens properties for better generalizability by diminishing the data bias in the network. We demonstrate the effectiveness of WorldGen by presenting an evaluation on deep optical flow. We hope such a tool can open doors for future research in a myriad of domains related to robotics and computer vision by reducing manual labor and the cost of acquiring rich and high-quality data.

Chethan M. Parameshwara, Gokul Hari, Cornelia Fermüller et al.

Current deep neural network approaches for camera pose estimation rely on scene structure for 3D motion estimation, but this decreases the robustness and thereby makes cross-dataset generalization difficult. In contrast, classical approaches to structure from motion estimate 3D motion utilizing optical flow and then compute depth. Their accuracy, however, depends strongly on the quality of the optical flow. To avoid this issue, direct methods have been proposed, which separate 3D motion from depth estimation but compute 3D motion using only image gradients in the form of normal flow. In this paper, we introduce a network NFlowNet, for normal flow estimation which is used to enforce robust and direct constraints. In particular, normal flow is used to estimate relative camera pose based on the cheirality (depth positivity) constraint. We achieve this by formulating the optimization problem as a differentiable cheirality layer, which allows for end-to-end learning of camera pose. We perform extensive qualitative and quantitative evaluation of the proposed DiffPoseNet's sensitivity to noise and its generalization across datasets. We compare our approach to existing state-of-the-art methods on KITTI, TartanAir, and TUM-RGBD datasets.

Amir-Hossein Shahidzadeh, Seong Jong Yoo, Pavan Mantripragada et al.

Tactile exploration plays a crucial role in understanding object structures for fundamental robotics tasks such as grasping and manipulation. However, efficiently exploring such objects using tactile sensors is challenging, primarily due to the large-scale unknown environments and limited sensing coverage of these sensors. To this end, we present AcTExplore, an active tactile exploration method driven by reinforcement learning for object reconstruction at scales that automatically explores the object surfaces in a limited number of steps. Through sufficient exploration, our algorithm incrementally collects tactile data and reconstructs 3D shapes of the objects as well, which can serve as a representation for higher-level downstream tasks. Our method achieves an average of 95.97% IoU coverage on unseen YCB objects while just being trained on primitive shapes. Project Webpage: https://prg.cs.umd.edu/AcTExplore

Levi Burner, Nitin J. Sanket, Cornelia Fermüller et al.

Distance estimation from vision is fundamental for a myriad of robotic applications such as navigation, manipulation, and planning. Inspired by the mammal's visual system, which gazes at specific objects, we develop two novel constraints relating time-to-contact, acceleration, and distance that we call the $τ$-constraint and $Φ$-constraint. They allow an active (moving) camera to estimate depth efficiently and accurately while using only a small portion of the image. The constraints are applicable to range sensing, sensor fusion, and visual servoing. We successfully validate the proposed constraints with two experiments. The first applies both constraints in a trajectory estimation task with a monocular camera and an Inertial Measurement Unit (IMU). Our methods achieve 30-70% less average trajectory error while running 25$\times$ and 6.2$\times$ faster than the popular Visual-Inertial Odometry methods VINS-Mono and ROVIO respectively. The second experiment demonstrates that when the constraints are used for feedback with efference copies the resulting closed loop system's eigenvalues are invariant to scaling of the applied control signal. We believe these results indicate the $τ$ and $Φ$ constraint's potential as the basis of robust and efficient algorithms for a multitude of robotic applications.

Dehao Yuan, Furong Huang, Cornelia Fermüller et al.

We propose Hyper-Dimensional Function Encoding (HDFE). Given samples of a continuous object (e.g. a function), HDFE produces an explicit vector representation of the given object, invariant to the sample distribution and density. Sample distribution and density invariance enables HDFE to consistently encode continuous objects regardless of their sampling, and therefore allows neural networks to receive continuous objects as inputs for machine learning tasks, such as classification and regression. Besides, HDFE does not require any training and is proved to map the object into an organized embedding space, which facilitates the training of the downstream tasks. In addition, the encoding is decodable, which enables neural networks to regress continuous objects by regressing their encodings. Therefore, HDFE serves as an interface for processing continuous objects. We apply HDFE to function-to-function mapping, where vanilla HDFE achieves competitive performance as the state-of-the-art algorithm. We apply HDFE to point cloud surface normal estimation, where a simple replacement from PointNet to HDFE leads to immediate 12% and 15% error reductions in two benchmarks. In addition, by integrating HDFE into the PointNet-based SOTA network, we improve the SOTA baseline by 2.5% and 1.7% in the same benchmarks.

Eadom Dessalene, Michael Maynord, Cornelia Fermüller et al.

We introduce LEAP (illustrated in Figure 1), a novel method for generating video-grounded action programs through use of a Large Language Model (LLM). These action programs represent the motoric, perceptual, and structural aspects of action, and consist of sub-actions, pre- and post-conditions, and control flows. LEAP's action programs are centered on egocentric video and employ recent developments in LLMs both as a source for program knowledge and as an aggregator and assessor of multimodal video information. We apply LEAP over a majority (87\%) of the training set of the EPIC Kitchens dataset, and release the resulting action programs as a publicly available dataset here (https://drive.google.com/drive/folders/1Cpkw_TI1IIxXdzor0pOXG3rWJWuKU5Ex?usp=drive_link). We employ LEAP as a secondary source of supervision, using its action programs in a loss term applied to action recognition and anticipation networks. We demonstrate sizable improvements in performance in both tasks due to training with the LEAP dataset. Our method achieves 1st place on the EPIC Kitchens Action Recognition leaderboard as of November 17 among the networks restricted to RGB-input (see Supplementary Materials).

Matthew Evanusa, Cornelia Fermüller, Yiannis Aloimonos

Artificial Neural Networks has struggled to devise a way to incorporate working memory into neural networks. While the ``long term'' memory can be seen as the learned weights, the working memory consists likely more of dynamical activity, that is missing from feed-forward models. Current state of the art models such as transformers tend to ``solve'' this by ignoring working memory entirely and simply process the sequence as an entire piece of data; however this means the network cannot process the sequence in an online fashion, and leads to an immense explosion in memory requirements. Here, inspired by a combination of controls, reservoir computing, deep learning, and recurrent neural networks, we offer an alternative paradigm that combines the strength of recurrent networks, with the pattern matching capability of feed-forward neural networks, which we call the \textit{Maelstrom Networks} paradigm. This paradigm leaves the recurrent component - the \textit{Maelstrom} - unlearned, and offloads the learning to a powerful feed-forward network. This allows the network to leverage the strength of feed-forward training without unrolling the network, and allows for the memory to be implemented in new neuromorphic hardware. It endows a neural network with a sequential memory that takes advantage of the inductive bias that data is organized causally in the temporal domain, and imbues the network with a state that represents the agent's ``self'', moving through the environment. This could also lead the way to continual learning, with the network modularized and ``'protected'' from overwrites that come with new data. In addition to aiding in solving these performance problems that plague current non-temporal deep networks, this also could finally lead towards endowing artificial networks with a sense of ``self''.

Jean-Marc Fellous, Gert Cauwenberghs, Cornelia Fermüller et al.

Neuroscience and Artificial Intelligence (AI) have made significant progress in the past few years but have only been loosely inter-connected. Based on a workshop held in August 2025, we identify current and future areas of synergism between these two fields. We focus on the subareas of embodiment, language and communication, robotics, learning in humans and machines and Neuromorphic engineering to take stock of the progress made so far, and possible promising new future avenues. Overall, we advocate for the development of NeuroAI, a type of Neuroscience-informed Artificial Intelligence that, we argue, has the potential for significantly improving the scope and efficiency of AI algorithms while simultaneously changing the way we understand biological neural computations. We include personal statements from several leading researchers on their diverse views of NeuroAI. Two Strength-Weakness-Opportunities-Threat (SWOT) analyses by researchers and trainees are appended that describe the benefits and risks offered by NeuroAI.

Dehao Yuan, Cornelia Fermüller

This paper presents a real-time, asynchronous, event-based normal flow estimator. It follows the same algorithm as Learning Normal Flow Directly From Event Neighborhoods, but with a more optimized implementation. The original method treats event slices as 3D point clouds, encodes each event's local geometry into a fixed-length vector, and uses a multi-layer perceptron to predict normal flow. It constructs representations by multiplying an adjacency matrix with a feature matrix, resulting in quadratic time complexity with respect to the number of events. In contrast, we leverage the fact that event coordinates are integers and reformulate the representation step as a pooling operation. This achieves the same effect as the adjacency matrix but with much lower computational cost. As a result, our method supports real-time normal flow prediction on event cameras. Our estimator uses 1 GB of CUDA memory and runs at 4 million normal flows per second on an RTX 3070, or 6 million per second on an RTX A5000. We release the CUDA implementation along with a Python interface at https://github.com/dhyuan99/VecKM_flow_cpp.

Jiayi Wu, Xiaomin Lin, Shahriar Negahdaripour et al.

Tasks such as autonomous navigation, 3D reconstruction, and object recognition near the water surfaces are crucial in marine robotics applications. However, challenges arise due to dynamic disturbances, e.g., light reflections and refraction from the random air-water interface, irregular liquid flow, and similar factors, which can lead to potential failures in perception and navigation systems. Traditional computer vision algorithms struggle to differentiate between real and virtual image regions, significantly complicating tasks. A virtual image region is an apparent representation formed by the redirection of light rays, typically through reflection or refraction, creating the illusion of an object's presence without its actual physical location. This work proposes a novel approach for segmentation on real and virtual image regions, exploiting synthetic images combined with domain-invariant information, a Motion Entropy Kernel, and Epipolar Geometric Consistency. Our segmentation network does not need to be re-trained if the domain changes. We show this by deploying the same segmentation network in two different domains: simulation and the real world. By creating realistic synthetic images that mimic the complexities of the water surface, we provide fine-grained training data for our network (MARVIS) to discern between real and virtual images effectively. By motion & geometry-aware design choices and through comprehensive experimental analysis, we achieve state-of-the-art real-virtual image segmentation performance in unseen real world domain, achieving an IoU over 78% and a F1-Score over 86% while ensuring a small computational footprint. MARVIS offers over 43 FPS (8 FPS) inference rates on a single GPU (CPU core). Our code and dataset are available here https://github.com/jiayi-wu-umd/MARVIS.

Chethan M. Parameshwara, Nitin J. Sanket, Chahat Deep Singh et al.

Segmentation of moving objects in dynamic scenes is a key process in scene understanding for navigation tasks. Classical cameras suffer from motion blur in such scenarios rendering them effete. On the contrary, event cameras, because of their high temporal resolution and lack of motion blur, are tailor-made for this problem. We present an approach for monocular multi-motion segmentation, which combines bottom-up feature tracking and top-down motion compensation into a unified pipeline, which is the first of its kind to our knowledge. Using the events within a time-interval, our method segments the scene into multiple motions by splitting and merging. We further speed up our method by using the concept of motion propagation and cluster keyslices. The approach was successfully evaluated on both challenging real-world and synthetic scenarios from the EV-IMO, EED, and MOD datasets and outperformed the state-of-the-art detection rate by 12\%, achieving a new state-of-the-art average detection rate of 81.06%, 94.2% and 82.35% on the aforementioned datasets. To enable further research and systematic evaluation of multi-motion segmentation, we present and open-source a new dataset/benchmark called MOD++, which includes challenging sequences and extensive data stratification in-terms of camera and object motion, velocity magnitudes, direction, and rotational speeds.

Dehao Yuan, Levi Burner, Jiayi Wu et al.

Event-based motion field estimation is an important task. However, current optical flow methods face challenges: learning-based approaches, often frame-based and relying on CNNs, lack cross-domain transferability, while model-based methods, though more robust, are less accurate. To address the limitations of optical flow estimation, recent works have focused on normal flow, which can be more reliably measured in regions with limited texture or strong edges. However, existing normal flow estimators are predominantly model-based and suffer from high errors. In this paper, we propose a novel supervised point-based method for normal flow estimation that overcomes the limitations of existing event learning-based approaches. Using a local point cloud encoder, our method directly estimates per-event normal flow from raw events, offering multiple unique advantages: 1) It produces temporally and spatially sharp predictions. 2) It supports more diverse data augmentation, such as random rotation, to improve robustness across various domains. 3) It naturally supports uncertainty quantification via ensemble inference, which benefits downstream tasks. 4) It enables training and inference on undistorted data in normalized camera coordinates, improving transferability across cameras. Extensive experiments demonstrate our method achieves better and more consistent performance than state-of-the-art methods when transferred across different datasets. Leveraging this transferability, we train our model on the union of datasets and release it for public use. Finally, we introduce an egomotion solver based on a maximum-margin problem that uses normal flow and IMU to achieve strong performance in challenging scenarios.

Dehao Yuan, Cornelia Fermüller, Tahseen Rabbani et al.

We propose VecKM, a local point cloud geometry encoder that is descriptive and efficient to compute. VecKM leverages a unique approach by vectorizing a kernel mixture to represent the local point cloud. Such representation's descriptiveness is supported by two theorems that validate its ability to reconstruct and preserve the similarity of the local shape. Unlike existing encoders downsampling the local point cloud, VecKM constructs the local geometry encoding using all neighboring points, producing a more descriptive encoding. Moreover, VecKM is efficient to compute and scalable to large point cloud inputs: VecKM reduces the memory cost from $(n^2+nKd)$ to $(nd+np)$; and reduces the major runtime cost from computing $nK$ MLPs to $n$ MLPs, where $n$ is the size of the point cloud, $K$ is the neighborhood size, $d$ is the encoding dimension, and $p$ is a marginal factor. The efficiency is due to VecKM's unique factorizable property that eliminates the need of explicitly grouping points into neighbors. In the normal estimation task, VecKM demonstrates not only 100x faster inference speed but also highest accuracy and strongest robustness. In classification and segmentation tasks, integrating VecKM as a preprocessing module achieves consistently better performance than the PointNet, PointNet++, and point transformer baselines, and runs consistently faster by up to 10 times.

Minghui Liu, Tahseen Rabbani, Tony O'Halloran et al.

Transformer-based large language models (LLMs) use the key-value (KV) cache to significantly accelerate inference by storing the key and value embeddings of past tokens. However, this cache consumes significant GPU memory. In this work, we introduce HashEvict, an algorithm that uses locality-sensitive hashing (LSH) to compress the KV cache. HashEvict quickly locates tokens in the cache that are cosine dissimilar to the current query token. This is achieved by computing the Hamming distance between binarized Gaussian projections of the current token query and cached token keys, with a projection length much smaller than the embedding dimension. We maintain a lightweight binary structure in GPU memory to facilitate these calculations. Unlike existing compression strategies that compute attention to determine token retention, HashEvict makes these decisions pre-attention, thereby reducing computational costs. Additionally, HashEvict is dynamic - at every decoding step, the key and value of the current token replace the embeddings of a token expected to produce the lowest attention score. We demonstrate that HashEvict can compress the KV cache by 30%-70% while maintaining high performance across reasoning, multiple-choice, long-context retrieval and summarization tasks.

Amir-Hossein Shahidzadeh, Jiyue Zhu, Kezhou Chen et al.

Robust object pose estimation is essential for manipulation and interaction tasks in robotics, particularly in scenarios where visual data is limited or sensitive to lighting, occlusions, and appearances. Tactile sensors often offer limited and local contact information, making it challenging to reconstruct the pose from partial data. Our approach uses sensorimotor exploration to actively control a robot hand to interact with the object. We train with Reinforcement Learning (RL) to explore and collect tactile data. The collected 3D point clouds are used to iteratively refine the object's shape and pose. In our setup, one hand holds the object steady while the other performs active exploration. We show that our method can actively explore an object's surface to identify critical pose features without prior knowledge of the object's geometry. Supplementary material and more demonstrations will be provided at https://amirshahid.github.io/BimanualTactilePose .

Zhiyuan Hua, Dehao Yuan, Cornelia Fermüller

This paper introduces a robust framework for motion segmentation and egomotion estimation using event-based normal flow, tailored specifically for neuromorphic vision sensors. In contrast to traditional methods that rely heavily on optical flow or explicit depth estimation, our approach exploits the sparse, high-temporal-resolution event data and incorporates geometric constraints between normal flow, scene structure, and inertial measurements. The proposed optimization-based pipeline iteratively performs event over-segmentation, isolates independently moving objects via residual analysis, and refines segmentations using hierarchical clustering informed by motion similarity and temporal consistency. Experimental results on the EVIMO2v2 dataset validate that our method achieves accurate segmentation and translational motion estimation without requiring full optical flow computation. This approach demonstrates significant advantages at object boundaries and offers considerable potential for scalable, real-time robotic and navigation applications.

Snehesh Shrestha, Yantian Zha, Saketh Banagiri et al.

Recent advances in multimodal Human-Robot Interaction (HRI) datasets emphasize the integration of speech and gestures, allowing robots to absorb explicit knowledge and tacit understanding. However, existing datasets primarily focus on elementary tasks like object pointing and pushing, limiting their applicability to complex domains. They prioritize simpler human command data but place less emphasis on training robots to correctly interpret tasks and respond appropriately. To address these gaps, we present the NatSGLD dataset, which was collected using a Wizard of Oz (WoZ) method, where participants interacted with a robot they believed to be autonomous. NatSGLD records humans' multimodal commands (speech and gestures), each paired with a demonstration trajectory and a Linear Temporal Logic (LTL) formula that provides a ground-truth interpretation of the commanded tasks. This dataset serves as a foundational resource for research at the intersection of HRI and machine learning. By providing multimodal inputs and detailed annotations, NatSGLD enables exploration in areas such as multimodal instruction following, plan recognition, and human-advisable reinforcement learning from demonstrations. We release the dataset and code under the MIT License at https://www.snehesh.com/natsgld/ to support future HRI research.

Seong Jong Yoo, Snehesh Shrestha, Irina Muresanu et al.

Musicians delicately control their bodies to generate music. Sometimes, their motions are too subtle to be captured by the human eye. To analyze how they move to produce the music, we need to estimate precise 4D human pose (3D pose over time). However, current state-of-the-art (SoTA) visual pose estimation algorithms struggle to produce accurate monocular 4D poses because of occlusions, partial views, and human-object interactions. They are limited by the viewing angle, pixel density, and sampling rate of the cameras and fail to estimate fast and subtle movements, such as in the musical effect of vibrato. We leverage the direct causal relationship between the music produced and the human motions creating them to address these challenges. We propose VioPose: a novel multimodal network that hierarchically estimates dynamics. High-level features are cascaded to low-level features and integrated into Bayesian updates. Our architecture is shown to produce accurate pose sequences, facilitating precise motion analysis, and outperforms SoTA. As part of this work, we collected the largest and the most diverse calibrated violin-playing dataset, including video, sound, and 3D motion capture poses. Code and dataset can be found in our project page \url{https://sj-yoo.info/viopose/}.

Siyuan Peng, Kate Ladenheim, Snehesh Shrestha et al.

This paper introduces the concept of a design tool for artistic performances based on attribute descriptions. To do so, we used a specific performance of falling actions. The platform integrates a novel machine-learning (ML) model with an interactive interface to generate and visualize artistic movements. Our approach's core is a cyclic Attribute-Conditioned Variational Autoencoder (AC-VAE) model developed to address the challenge of capturing and generating realistic 3D human body motions from motion capture (MoCap) data. We created a unique dataset focused on the dynamics of falling movements, characterized by a new ontology that divides motion into three distinct phases: Impact, Glitch, and Fall. The ML model's innovation lies in its ability to learn these phases separately. It is achieved by applying comprehensive data augmentation techniques and an initial pose loss function to generate natural and plausible motion. Our web-based interface provides an intuitive platform for artists to engage with this technology, offering fine-grained control over motion attributes and interactive visualization tools, including a 360-degree view and a dynamic timeline for playback manipulation. Our research paves the way for a future where technology amplifies the creative potential of human expression, making sophisticated motion generation accessible to a wider artistic community.

Chahat Deep Singh, Nitin J. Sanket, Chethan M. Parameshwara et al.

Recent advances in object segmentation have demonstrated that deep neural networks excel at object segmentation for specific classes in color and depth images. However, their performance is dictated by the number of classes and objects used for training, thereby hindering generalization to never seen objects or zero-shot samples. To exacerbate the problem further, object segmentation using image frames rely on recognition and pattern matching cues. Instead, we utilize the 'active' nature of a robot and their ability to 'interact' with the environment to induce additional geometric constraints for segmenting zero-shot samples. In this paper, we present the first framework to segment unknown objects in a cluttered scene by repeatedly 'nudging' at the objects and moving them to obtain additional motion cues at every step using only a monochrome monocular camera. We call our framework NudgeSeg. These motion cues are used to refine the segmentation masks. We successfully test our approach to segment novel objects in various cluttered scenes and provide an extensive study with image and motion segmentation methods. We show an impressive average detection rate of over 86% on zero-shot objects.

Nitin J. Sanket, Chahat Deep Singh, Chethan M. Parameshwara et al.

The rapid rise of accessibility of unmanned aerial vehicles or drones pose a threat to general security and confidentiality. Most of the commercially available or custom-built drones are multi-rotors and are comprised of multiple propellers. Since these propellers rotate at a high-speed, they are generally the fastest moving parts of an image and cannot be directly "seen" by a classical camera without severe motion blur. We utilize a class of sensors that are particularly suitable for such scenarios called event cameras, which have a high temporal resolution, low-latency, and high dynamic range. In this paper, we model the geometry of a propeller and use it to generate simulated events which are used to train a deep neural network called EVPropNet to detect propellers from the data of an event camera. EVPropNet directly transfers to the real world without any fine-tuning or retraining. We present two applications of our network: (a) tracking and following an unmarked drone and (b) landing on a near-hover drone. We successfully evaluate and demonstrate the proposed approach in many real-world experiments with different propeller shapes and sizes. Our network can detect propellers at a rate of 85.1% even when 60% of the propeller is occluded and can run at upto 35Hz on a 2W power budget. To our knowledge, this is the first deep learning-based solution for detecting propellers (to detect drones). Finally, our applications also show an impressive success rate of 92% and 90% for the tracking and landing tasks respectively.

Chethan M. Parameshwara, Simin Li, Cornelia Fermüller et al.

Spiking Neural Networks (SNN) are the so-called third generation of neural networks which attempt to more closely match the functioning of the biological brain. They inherently encode temporal data, allowing for training with less energy usage and can be extremely energy efficient when coded on neuromorphic hardware. In addition, they are well suited for tasks involving event-based sensors, which match the event-based nature of the SNN. However, SNNs have not been as effectively applied to real-world, large-scale tasks as standard Artificial Neural Networks (ANNs) due to the algorithmic and training complexity. To exacerbate the situation further, the input representation is unconventional and requires careful analysis and deep understanding. In this paper, we propose \textit{SpikeMS}, the first deep encoder-decoder SNN architecture for the real-world large-scale problem of motion segmentation using the event-based DVS camera as input. To accomplish this, we introduce a novel spatio-temporal loss formulation that includes both spike counts and classification labels in conjunction with the use of new techniques for SNN backpropagation. In addition, we show that \textit{SpikeMS} is capable of \textit{incremental predictions}, or predictions from smaller amounts of test data than it is trained on. This is invaluable for providing outputs even with partial input data for low-latency applications and those requiring fast predictions. We evaluated \textit{SpikeMS} on challenging synthetic and real-world sequences from EV-IMO, EED and MOD datasets and achieving results on a par with a comparable ANN method, but using potentially 50 times less power.

Nitin J. Sanket, Chahat Deep Singh, Varun Asthana et al.

Morphable design and depth-based visual control are two upcoming trends leading to advancements in the field of quadrotor autonomy. Stereo-cameras have struck the perfect balance of weight and accuracy of depth estimation but suffer from the problem of depth range being limited and dictated by the baseline chosen at design time. In this paper, we present a framework for quadrotor navigation based on a stereo camera system whose baseline can be adapted on-the-fly. We present a method to calibrate the system at a small number of discrete baselines and interpolate the parameters for the entire baseline range. We present an extensive theoretical analysis of calibration and synchronization errors. We showcase three different applications of such a system for quadrotor navigation: (a) flying through a forest, (b) flying through an unknown shaped/location static/dynamic gap, and (c) accurate 3D pose detection of an independently moving object. We show that our variable baseline system is more accurate and robust in all three scenarios. To our knowledge, this is the first work that applies the concept of morphable design to achieve a variable baseline stereo vision system on a quadrotor.

Kanishka Ganguly, Behzad Sadrfaridpour, Pavan Mantripragada et al.

Grasping and manipulating a wide variety of objects is a fundamental skill that would determine the success and wide spread adaptation of robots in homes. Several end-effector designs for robust manipulation have been proposed but they mostly work when provided with prior information about the objects or equipped with external sensors for estimating object shape or size. Such approaches are limited to many-shot or unknown objects and are prone to estimation errors from external estimation systems. We propose an approach to grasp and manipulate previously unseen or zero-shot objects: the objects without any prior of their shape, size, material and weight properties, using only feedback from tactile sensors which is contrary to the state-of-the-art. Such an approach provides robust manipulation of objects either when the object model is not known or when it is estimated incorrectly from an external system. Our approach is inspired by the ideology of how animals or humans manipulate objects, i.e., by using feedback from their skin. Our grasping and manipulation revolves around the simple notion that objects slip if not grasped stably. This slippage can be detected and counteracted for a robust grasp that is agnostic to the type, shape, size, material and weight of the object. At the crux of our approach is a novel tactile feedback based controller that detects and compensates for slip during grasp. We successfully evaluate and demonstrate our proposed approach on many real world experiments using the Shadow Dexterous Hand equipped with BioTac SP tactile sensors for different object shapes, sizes, weights and materials. We obtain an overall success rate of 73.5%

Matthew Evanusa, Snehesh Shrestha, Michelle Girvan et al.

In many real-world applications, fully-differentiable RNNs such as LSTMs and GRUs have been widely deployed to solve time series learning tasks. These networks train via Backpropagation Through Time, which can work well in practice but involves a biologically unrealistic unrolling of the network in time for gradient updates, are computationally expensive, and can be hard to tune. A second paradigm, Reservoir Computing, keeps the recurrent weight matrix fixed and random. Here, we propose a novel hybrid network, which we call Hybrid Backpropagation Parallel Echo State Network (HBP-ESN) which combines the effectiveness of learning random temporal features of reservoirs with the readout power of a deep neural network with batch normalization. We demonstrate that our new network outperforms LSTMs and GRUs, including multi-layer "deep" versions of these networks, on two complex real-world multi-dimensional time series datasets: gesture recognition using skeleton keypoints from ChaLearn, and the DEAP dataset for emotion recognition from EEG measurements. We show also that the inclusion of a novel meta-ring structure, which we call HBP-ESN M-Ring, achieves similar performance to one large reservoir while decreasing the memory required by an order of magnitude. We thus offer this new hybrid reservoir deep learning paradigm as a new alternative direction for RNN learning of temporal or sequential data.

Matthew Evanusa, Cornelia Fermüller, Yiannis Aloimonos

Deep Reservoir Computing has emerged as a new paradigm for deep learning, which is based around the reservoir computing principle of maintaining random pools of neurons combined with hierarchical deep learning. The reservoir paradigm reflects and respects the high degree of recurrence in biological brains, and the role that neuronal dynamics play in learning. However, one issue hampering deep reservoir network development is that one cannot backpropagate through the reservoir layers. Recent deep reservoir architectures do not learn hidden or hierarchical representations in the same manner as deep artificial neural networks, but rather concatenate all hidden reservoirs together to perform traditional regression. Here we present a novel Deep Reservoir Network for time series prediction and classification that learns through the non-differentiable hidden reservoir layers using a biologically-inspired backpropagation alternative called Direct Feedback Alignment, which resembles global dopamine signal broadcasting in the brain. We demonstrate its efficacy on two real world multidimensional time series datasets.

Nitin J. Sanket, Chahat Deep Singh, Cornelia Fermüller et al.

Odometry on aerial robots has to be of low latency and high robustness whilst also respecting the Size, Weight, Area and Power (SWAP) constraints as demanded by the size of the robot. A combination of visual sensors coupled with Inertial Measurement Units (IMUs) has proven to be the best combination to obtain robust and low latency odometry on resource-constrained aerial robots. Recently, deep learning approaches for Visual Inertial fusion have gained momentum due to their high accuracy and robustness. However, the remarkable advantages of these techniques are their inherent scalability (adaptation to different sized aerial robots) and unification (same method works on different sized aerial robots) by utilizing compression methods and hardware acceleration, which have been lacking from previous approaches. To this end, we present a deep learning approach for visual translation estimation and loosely fuse it with an Inertial sensor for full 6DoF odometry estimation. We also present a detailed benchmark comparing different architectures, loss functions and compression methods to enable scalability. We evaluate our network on the MSCOCO dataset and evaluate the VI fusion on multiple real-flight trajectories.

Nitin J. Sanket, Chethan M. Parameshwara, Chahat Deep Singh et al.

Dynamic obstacle avoidance on quadrotors requires low latency. A class of sensors that are particularly suitable for such scenarios are event cameras. In this paper, we present a deep learning -- based solution for dodging multiple dynamic obstacles on a quadrotor with a single event camera and on-board computation. Our approach uses a series of shallow neural networks for estimating both the ego-motion and the motion of independently moving objects. The networks are trained in simulation and directly transfer to the real world without any fine-tuning or retraining. We successfully evaluate and demonstrate the proposed approach in many real-world experiments with obstacles of different shapes and sizes, achieving an overall success rate of 70% including objects of unknown shape and a low light testing scenario. To our knowledge, this is the first deep learning -- based solution to the problem of dynamic obstacle avoidance using event cameras on a quadrotor. Finally, we also extend our work to the pursuit task by merely reversing the control policy, proving that our navigation stack can cater to different scenarios.

Chengxi Ye, Matthew Evanusa, Hua He et al.

Convolution is a central operation in Convolutional Neural Networks (CNNs), which applies a kernel to overlapping regions shifted across the image. However, because of the strong correlations in real-world image data, convolutional kernels are in effect re-learning redundant data. In this work, we show that this redundancy has made neural network training challenging, and propose network deconvolution, a procedure which optimally removes pixel-wise and channel-wise correlations before the data is fed into each layer. Network deconvolution can be efficiently calculated at a fraction of the computational cost of a convolution layer. We also show that the deconvolution filters in the first layer of the network resemble the center-surround structure found in biological neurons in the visual regions of the brain. Filtering with such kernels results in a sparse representation, a desired property that has been missing in the training of neural networks. Learning from the sparse representation promotes faster convergence and superior results without the use of batch normalization. We apply our network deconvolution operation to 10 modern neural network models by replacing batch normalization within each. Extensive experiments show that the network deconvolution operation is able to deliver performance improvement in all cases on the CIFAR-10, CIFAR-100, MNIST, Fashion-MNIST, Cityscapes, and ImageNet datasets.

Kanishka Ganguly, Behzad Sadrfaridpour, Cornelia Fermüller et al.

Grasping objects requires tight integration between visual and tactile feedback. However, there is an inherent difference in the scale at which both these input modalities operate. It is thus necessary to be able to analyze tactile feedback in isolation in order to gain information about the surface the end-effector is operating on, such that more fine-grained features may be extracted from the surroundings. For tactile perception of the robot, inspired by the concept of the tactile flow in humans, we present the computational tactile flow to improve the analysis of the tactile feedback in robots using a Shadow Dexterous Hand. In the computational tactile flow model, given a sequence of pressure values from the tactile sensors, we define a virtual surface for the pressure values and define the tactile flow as the optical flow of this surface. We provide case studies that demonstrate how the computational tactile flow maps reveal information on the direction of motion and 3D structure of the surface, and feedback regarding the action being performed by the robot.

Chengxi Ye, Anton Mitrokhin, Cornelia Fermüller et al.

In this work we present a lightweight, unsupervised learning pipeline for \textit{dense} depth, optical flow and egomotion estimation from sparse event output of the Dynamic Vision Sensor (DVS). To tackle this low level vision task, we use a novel encoder-decoder neural network architecture - ECN. Our work is the first monocular pipeline that generates dense depth and optical flow from sparse event data only. The network works in self-supervised mode and has just 150k parameters. We evaluate our pipeline on the MVSEC self driving dataset and present results for depth, optical flow and and egomotion estimation. Due to the lightweight design, the inference part of the network runs at 250 FPS on a single GPU, making the pipeline ready for realtime robotics applications. Our experiments demonstrate significant improvements upon previous works that used deep learning on event data, as well as the ability of our pipeline to perform well during both day and night.

Chengxi Ye, Chinmaya Devaraj, Michael Maynord et al.

We introduce Evenly Cascaded convolutional Network (ECN), a neural network taking inspiration from the cascade algorithm of wavelet analysis. ECN employs two feature streams - a low-level and high-level steam. At each layer these streams interact, such that low-level features are modulated using advanced perspectives from the high-level stream. ECN is evenly structured through resizing feature map dimensions by a consistent ratio, which removes the burden of ad-hoc specification of feature map dimensions. ECN produces easily interpretable features maps, a result whose intuition can be understood in the context of scale-space theory. We demonstrate that ECN's design facilitates the training process through providing easily trainable shortcuts. We report new state-of-the-art results for small networks, without the need for additional treatment such as pruning or compression - a consequence of ECN's simple structure and direct training. A 6-layered ECN design with under 500k parameters achieves 95.24% and 78.99% accuracy on CIFAR-10 and CIFAR-100 datasets, respectively, outperforming the current state-of-the-art on small parameter networks, and a 3 million parameter ECN produces results competitive to the state-of-the-art.

Francisco Barranco, Cornelia Fermüller, Yiannis Aloimonos et al.

Conventional image motion based structure from motion methods first compute optical flow, then solve for the 3D motion parameters based on the epipolar constraint, and finally recover the 3D geometry of the scene. However, errors in optical flow due to regularization can lead to large errors in 3D motion and structure. This paper investigates whether performance and consistency can be improved by avoiding optical flow estimation in the early stages of the structure from motion pipeline, and it proposes a new direct method based on image gradients (normal flow) only. The main idea lies in a reformulation of the positive-depth constraint, which allows the use of well-known minimization techniques to solve for 3D motion. The 3D motion estimate is then refined and structure estimated adding a regularization based on depth. Experimental comparisons on standard synthetic datasets and the real-world driving benchmark dataset KITTI using three different optic flow algorithms show that the method achieves better accuracy in all but one case. Furthermore, it outperforms existing normal flow based 3D motion estimation techniques. Finally, the recovered 3D geometry is shown to be also very accurate.

Huai-Jen Liang, Nitin J. Sanket, Cornelia Fermüller et al.

Although cluttered indoor scenes have a lot of useful high-level semantic information which can be used for mapping and localization, most Visual Odometry (VO) algorithms rely on the usage of geometric features such as points, lines and planes. Lately, driven by this idea, the joint optimization of semantic labels and obtaining odometry has gained popularity in the robotics community. The joint optimization is good for accurate results but is generally very slow. At the same time, in the vision community, direct and sparse approaches for VO have stricken the right balance between speed and accuracy. We merge the successes of these two communities and present a way to incorporate semantic information in the form of visual saliency to Direct Sparse Odometry - a highly successful direct sparse VO algorithm. We also present a framework to filter the visual saliency based on scene parsing. Our framework, SalientDSO, relies on the widely successful deep learning based approaches for visual saliency and scene parsing which drives the feature selection for obtaining highly-accurate and robust VO even in the presence of as few as 40 point features per frame. We provide extensive quantitative evaluation of SalientDSO on the ICL-NUIM and TUM monoVO datasets and show that we outperform DSO and ORB-SLAM - two very popular state-of-the-art approaches in the literature. We also collect and publicly release a CVL-UMD dataset which contains two indoor cluttered sequences on which we show qualitative evaluations. To our knowledge this is the first paper to use visual saliency and scene parsing to drive the feature selection in direct VO.

Nitin J Sanket, Chahat Deep Singh, Kanishka Ganguly et al.

Although quadrotors, and aerial robots in general, are inherently active agents, their perceptual capabilities in literature so far have been mostly passive in nature. Researchers and practitioners today use traditional computer vision algorithms with the aim of building a representation of general applicability: a 3D reconstruction of the scene. Using this representation, planning tasks are constructed and accomplished to allow the quadrotor to demonstrate autonomous behavior. These methods are inefficient as they are not task driven and such methodologies are not utilized by flying insects and birds. Such agents have been solving the problem of navigation and complex control for ages without the need to build a 3D map and are highly task driven. In this paper, we propose this framework of bio-inspired perceptual design for quadrotors. We use this philosophy to design a minimalist sensori-motor framework for a quadrotor to fly though unknown gaps without a 3D reconstruction of the scene using only a monocular camera and onboard sensing. We successfully evaluate and demonstrate the proposed approach in many real-world experiments with different settings and window shapes, achieving a success rate of 85% at 2.5ms$^{-1}$ even with a minimum tolerance of just 5cm. To our knowledge, this is the first paper which addresses the problem of gap detection of an unknown shape and location with a monocular camera and onboard sensing.

Cornelia Fermüller, Fang Wang, Yezhou Yang et al.

Looking at a person's hands one often can tell what the person is going to do next, how his/her hands are moving and where they will be, because an actor's intentions shape his/her movement kinematics during action execution. Similarly, active systems with real-time constraints must not simply rely on passive video-segment classification, but they have to continuously update their estimates and predict future actions. In this paper, we study the prediction of dexterous actions. We recorded from subjects performing different manipulation actions on the same object, such as "squeezing", "flipping", "washing", "wiping" and "scratching" with a sponge. In psychophysical experiments, we evaluated human observers' skills in predicting actions from video sequences of different length, depicting the hand movement in the preparation and execution of actions before and after contact with the object. We then developed a recurrent neural network based method for action prediction using as input patches around the hand. We also used the same formalism to predict the forces on the finger tips using for training synchronized video and force data streams. Evaluations on two new datasets showed that our system closely matches human performance in the recognition task, and demonstrate the ability of our algorithm to predict what and how a dexterous action is performed.

Morimichi Nishigaki, Cornelia Fermüller

Contours are salient features for image description, but the detection and localization of boundary contours is still considered a challenging problem. This paper introduces a new tool for edge processing implementing the Gestaltism idea of edge grouping. This tool is a mid-level image operator, called the Torque operator, that is designed to help detect closed contours in images. The torque operator takes as input the raw image and creates an image map by computing from the image gradients within regions of multiple sizes a measure of how well the edges are aligned to form closed convex contours. Fundamental properties of the torque are explored and illustrated through examples. Then it is applied in pure bottom-up processing in a variety of applications, including edge detection, visual attention and segmentation and experimentally demonstrated a useful tool that can improve existing techniques. Finally, its extension as a more general grouping mechanism and application in object recognition is discussed.