Rahul Sukthankar

Vision Transformer (ViT) is emerging as the state-of-the-art architecture for image recognition. While recent studies suggest that ViTs are more robust than their convolutional counterparts, our experiments find that ViTs trained on ImageNet are overly reliant on local textures and fail to make adequate use of shape information. ViTs thus have difficulties generalizing to out-of-distribution, real-world data. To address this deficiency, we present a simple and effective architecture modification to ViT's input layer by adding discrete tokens produced by a vector-quantized encoder. Different from the standard continuous pixel tokens, discrete tokens are invariant under small perturbations and contain less information individually, which promote ViTs to learn global information that is invariant. Experimental results demonstrate that adding discrete representation on four architecture variants strengthens ViT robustness by up to 12% across seven ImageNet robustness benchmarks while maintaining the performance on ImageNet.

Advances in the state of the art for 3d human sensing are currently limited by the lack of visual datasets with 3d ground truth, including multiple people, in motion, operating in real-world environments, with complex illumination or occlusion, and potentially observed by a moving camera. Sophisticated scene understanding would require estimating human pose and shape as well as gestures, towards representations that ultimately combine useful metric and behavioral signals with free-viewpoint photo-realistic visualisation capabilities. To sustain progress, we build a large-scale photo-realistic dataset, Human-SPACE (HSPACE), of animated humans placed in complex synthetic indoor and outdoor environments. We combine a hundred diverse individuals of varying ages, gender, proportions, and ethnicity, with hundreds of motions and scenes, as well as parametric variations in body shape (for a total of 1,600 different humans), in order to generate an initial dataset of over 1 million frames. Human animations are obtained by fitting an expressive human body model, GHUM, to single scans of people, followed by novel re-targeting and positioning procedures that support the realistic animation of dressed humans, statistical variation of body proportions, and jointly consistent scene placement of multiple moving people. Assets are generated automatically, at scale, and are compatible with existing real time rendering and game engines. The dataset with evaluation server will be made available for research. Our large-scale analysis of the impact of synthetic data, in connection with real data and weak supervision, underlines the considerable potential for continuing quality improvements and limiting the sim-to-real gap, in this practical setting, in connection with increased model capacity.

We present deep neural network methodology to reconstruct the 3d pose and shape of people, given an input RGB image. We rely on a recently introduced, expressivefull body statistical 3d human model, GHUM, trained end-to-end, and learn to reconstruct its pose and shape state in a self-supervised regime. Central to our methodology, is a learning to learn and optimize approach, referred to as HUmanNeural Descent (HUND), which avoids both second-order differentiation when training the model parameters,and expensive state gradient descent in order to accurately minimize a semantic differentiable rendering loss at test time. Instead, we rely on novel recurrent stages to update the pose and shape parameters such that not only losses are minimized effectively, but the process is meta-regularized in order to ensure end-progress. HUND's symmetry between training and testing makes it the first 3d human sensing architecture to natively support different operating regimes including self-supervised ones. In diverse tests, we show that HUND achieves very competitive results in datasets like H3.6M and 3DPW, aswell as good quality 3d reconstructions for complex imagery collected in-the-wild.

We address the challenging problem of semi-supervised learning in the context of multiple visual interpretations of the world by finding consensus in a graph of neural networks. Each graph node is a scene interpretation layer, while each edge is a deep net that transforms one layer at one node into another from a different node. During the supervised phase edge networks are trained independently. During the next unsupervised stage edge nets are trained on the pseudo-ground truth provided by consensus among multiple paths that reach the nets' start and end nodes. These paths act as ensemble teachers for any given edge and strong consensus is used for high-confidence supervisory signal. The unsupervised learning process is repeated over several generations, in which each edge becomes a "student" and also part of different ensemble "teachers" for training other students. By optimizing such consensus between different paths, the graph reaches consistency and robustness over multiple interpretations and generations, in the face of unknown labels. We give theoretical justifications of the proposed idea and validate it on a large dataset. We show how prediction of different representations such as depth, semantic segmentation, surface normals and pose from RGB input could be effectively learned through self-supervised consensus in our graph. We also compare to state-of-the-art methods for multi-task and semi-supervised learning and show superior performance.

Monocular 3D human pose and shape estimation is challenging due to the many degrees of freedom of the human body and thedifficulty to acquire training data for large-scale supervised learning in complex visual scenes. In this paper we present practical semi-supervised and self-supervised models that support training and good generalization in real-world images and video. Our formulation is based on kinematic latent normalizing flow representations and dynamics, as well as differentiable, semantic body part alignment loss functions that support self-supervised learning. In extensive experiments using 3D motion capture datasets like CMU, Human3.6M, 3DPW, or AMASS, as well as image repositories like COCO, we show that the proposed methods outperform the state of the art, supporting the practical construction of an accurate family of models based on large-scale training with diverse and incompletely labeled image and video data.

A physical selfie stick extends the user's reach, enabling the acquisition of personal photos that include more of the background scene. Similarly, a quadcopter can capture photos from vantage points unattainable by the user; but teleoperating a quadcopter to good viewpoints is a difficult task. This paper presents a natural interface for quadcopter photography, the SelfieDroneStick that allows the user to guide the quadcopter to the optimal vantage point based on the phone's sensors. Users specify the composition of their desired long-range selfies using their smartphone, and the quadcopter autonomously flies to a sequence of vantage points from where the desired shots can be taken. The robot controller is trained from a combination of real-world images and simulated flight data. This paper describes two key innovations required to deploy deep reinforcement learning models on a real robot: 1) an abstract state representation for transferring learning from simulation to the hardware platform, and 2) reward shaping and staging paradigms for training the controller. Both of these improvements were found to be essential in learning a robot controller from simulation that transfers successfully to the real robot.

Object detection models based on convolutional neural networks (CNNs) demonstrate impressive performance when trained on large-scale labeled datasets. While a generic object detector trained on such a dataset performs adequately in applications where the input data is similar to user photographs, the detector performs poorly on small objects, particularly ones with limited training data or imaged from uncommon viewpoints. Also, a specific room will have many objects that are missed by standard object detectors, frustrating a robot that continually operates in the same indoor environment. This paper describes a system for rapidly creating customized object detectors. Data is collected from a quadcopter that is teleoperated with an interactive interface. Once an object is selected, the quadcopter autonomously photographs the object from multiple viewpoints to %create training data that is used by DUNet (Dense Upscaled Net), collect data to train DUNet (Dense Upscaled Network), our proposed model for learning customized object detectors from scratch given limited data. Our experiments compare the performance of learning models from scratch with DUNet vs.\ fine tuning existing state of the art object detectors, both on our indoor robotics domain and on standard datasets.

State-of-the-art methods for video action recognition commonly use an ensemble of two networks: the spatial stream, which takes RGB frames as input, and the temporal stream, which takes optical flow as input. In recent work, both of these streams consist of 3D Convolutional Neural Networks, which apply spatiotemporal filters to the video clip before performing classification. Conceptually, the temporal filters should allow the spatial stream to learn motion representations, making the temporal stream redundant. However, we still see significant benefits in action recognition performance by including an entirely separate temporal stream, indicating that the spatial stream is "missing" some of the signal captured by the temporal stream. In this work, we first investigate whether motion representations are indeed missing in the spatial stream of 3D CNNs. Second, we demonstrate that these motion representations can be improved by distillation, by tuning the spatial stream to predict the outputs of the temporal stream, effectively combining both models into a single stream. Finally, we show that our Distilled 3D Network (D3D) achieves performance on par with two-stream approaches, using only a single model and with no need to compute optical flow.

Solving tasks with sparse rewards is a main challenge in reinforcement learning. While hierarchical controllers are an intuitive approach to this problem, current methods often require manual reward shaping, alternating training phases, or manually defined sub tasks. We introduce modulated policy hierarchies (MPH), that can learn end-to-end to solve tasks from sparse rewards. To achieve this, we study different modulation signals and exploration for hierarchical controllers. Specifically, we find that communicating via bit-vectors is more efficient than selecting one out of multiple skills, as it enables mixing between them. To facilitate exploration, MPH uses its different time scales for temporally extended intrinsic motivation at each level of the hierarchy. We evaluate MPH on the robotics tasks of pushing and sparse block stacking, where it outperforms recent baselines.

We consider the problem of retrieving objects from image data and learning to classify them into meaningful semantic categories with minimal supervision. To that end, we propose a fully differentiable unsupervised deep clustering approach to learn semantic classes in an end-to-end fashion without individual class labeling using only unlabeled object proposals. The key contributions of our work are 1) a kmeans clustering objective where the clusters are learned as parameters of the network and are represented as memory units, and 2) simultaneously building a feature representation, or embedding, while learning to cluster it. This approach shows promising results on two popular computer vision datasets: on CIFAR10 for clustering objects, and on the more complex and challenging Cityscapes dataset for semantically discovering classes which visually correspond to cars, people, and bicycles. Currently, the only supervision provided is segmentation objectness masks, but this method can be extended to use an unsupervised objectness-based object generation mechanism which will make the approach completely unsupervised.