Hyun Soo Park

Kristen Grauman, Andrew Westbury, Lorenzo Torresani et al. · cmu, gatech

We present Ego-Exo4D, a diverse, large-scale multimodal multiview video dataset and benchmark challenge. Ego-Exo4D centers around simultaneously-captured egocentric and exocentric video of skilled human activities (e.g., sports, music, dance, bike repair). 740 participants from 13 cities worldwide performed these activities in 123 different natural scene contexts, yielding long-form captures from 1 to 42 minutes each and 1,286 hours of video combined. The multimodal nature of the dataset is unprecedented: the video is accompanied by multichannel audio, eye gaze, 3D point clouds, camera poses, IMU, and multiple paired language descriptions -- including a novel "expert commentary" done by coaches and teachers and tailored to the skilled-activity domain. To push the frontier of first-person video understanding of skilled human activity, we also present a suite of benchmark tasks and their annotations, including fine-grained activity understanding, proficiency estimation, cross-view translation, and 3D hand/body pose. All resources are open sourced to fuel new research in the community. Project page: http://ego-exo4d-data.org/

Jae Shin Yoon, Duygu Ceylan, Tuanfeng Y. Wang et al.

Appearance of dressed humans undergoes a complex geometric transformation induced not only by the static pose but also by its dynamics, i.e., there exists a number of cloth geometric configurations given a pose depending on the way it has moved. Such appearance modeling conditioned on motion has been largely neglected in existing human rendering methods, resulting in rendering of physically implausible motion. A key challenge of learning the dynamics of the appearance lies in the requirement of a prohibitively large amount of observations. In this paper, we present a compact motion representation by enforcing equivariance -- a representation is expected to be transformed in the way that the pose is transformed. We model an equivariant encoder that can generate the generalizable representation from the spatial and temporal derivatives of the 3D body surface. This learned representation is decoded by a compositional multi-task decoder that renders high fidelity time-varying appearance. Our experiments show that our method can generate a temporally coherent video of dynamic humans for unseen body poses and novel views given a single view video.

Yasamin Jafarian, Tuanfeng Y. Wang, Duygu Ceylan et al.

Clothes undergo complex geometric deformations, which lead to appearance changes. To edit human videos in a physically plausible way, a texture map must take into account not only the garment transformation induced by the body movements and clothes fitting, but also its 3D fine-grained surface geometry. This poses, however, a new challenge of 3D reconstruction of dynamic clothes from an image or a video. In this paper, we show that it is possible to edit dressed human images and videos without 3D reconstruction. We estimate a geometry aware texture map between the garment region in an image and the texture space, a.k.a, UV map. Our UV map is designed to preserve isometry with respect to the underlying 3D surface by making use of the 3D surface normals predicted from the image. Our approach captures the underlying geometry of the garment in a self-supervised way, requiring no ground truth annotation of UV maps and can be readily extended to predict temporally coherent UV maps. We demonstrate that our method outperforms the state-of-the-art human UV map estimation approaches on both real and synthetic data.

Jinwook Huh, Jungseok Hong, Suveer Garg et al.

One of the challenging input settings for visual servoing is when the initial and goal camera views are far apart. Such settings are difficult because the wide baseline can cause drastic changes in object appearance and cause occlusions. This paper presents a novel self-supervised visual servoing method for wide baseline images which does not require 3D ground truth supervision. Existing approaches that regress absolute camera pose with respect to an object require 3D ground truth data of the object in the forms of 3D bounding boxes or meshes. We learn a coherent visual representation by leveraging a geometric property called 3D equivariance-the representation is transformed in a predictable way as a function of 3D transformation. To ensure that the feature-space is faithful to the underlying geodesic space, a geodesic preserving constraint is applied in conjunction with the equivariance. We design a Siamese network that can effectively enforce these two geometric properties without requiring 3D supervision. With the learned model, the relative transformation can be inferred simply by following the gradient in the learned space and used as feedback for closed-loop visual servoing. Our method is evaluated on objects from the YCB dataset, showing meaningful outperformance on a visual servoing task, or object alignment task with respect to state-of-the-art approaches that use 3D supervision. Ours yields more than 35% average distance error reduction and more than 90% success rate with 3cm error tolerance.

Jingfan Guo, Fabian Prada, Donglai Xiang et al.

Registering clothes from 4D scans with vertex-accurate correspondence is challenging, yet important for dynamic appearance modeling and physics parameter estimation from real-world data. However, previous methods either rely on texture information, which is not always reliable, or achieve only coarse-level alignment. In this work, we present a novel approach to enabling accurate surface registration of texture-less clothes with large deformation. Our key idea is to effectively leverage a shape prior learned from pre-captured clothing using diffusion models. We also propose a multi-stage guidance scheme based on learned functional maps, which stabilizes registration for large-scale deformation even when they vary significantly from training data. Using high-fidelity real captured clothes, our experiments show that the proposed approach based on diffusion models generalizes better than surface registration with VAE or PCA-based priors, outperforming both optimization-based and learning-based non-rigid registration methods for both interpolation and extrapolation tests.

Tien Do, Khiem Vuong, Hyun Soo Park

In this paper, we study a problem of egocentric scene understanding, i.e., predicting depths and surface normals from an egocentric image. Egocentric scene understanding poses unprecedented challenges: (1) due to large head movements, the images are taken from non-canonical viewpoints (i.e., tilted images) where existing models of geometry prediction do not apply; (2) dynamic foreground objects including hands constitute a large proportion of visual scenes. These challenges limit the performance of the existing models learned from large indoor datasets, such as ScanNet and NYUv2, which comprise predominantly upright images of static scenes. We present a multimodal spatial rectifier that stabilizes the egocentric images to a set of reference directions, which allows learning a coherent visual representation. Unlike unimodal spatial rectifier that often produces excessive perspective warp for egocentric images, the multimodal spatial rectifier learns from multiple directions that can minimize the impact of the perspective warp. To learn visual representations of the dynamic foreground objects, we present a new dataset called EDINA (Egocentric Depth on everyday INdoor Activities) that comprises more than 500K synchronized RGBD frames and gravity directions. Equipped with the multimodal spatial rectifier and the EDINA dataset, our proposed method on single-view depth and surface normal estimation significantly outperforms the baselines not only on our EDINA dataset, but also on other popular egocentric datasets, such as First Person Hand Action (FPHA) and EPIC-KITCHENS.

Zachary Chavis, Hyun Soo Park, Stephen J. Guy

Vision-Language Models (VLMs) have shown great success as foundational models for downstream vision and natural language applications in a variety of domains. However, these models are limited to reasoning over objects and actions currently visible on the image plane. We present a spatial extension to the VLM, which leverages spatially-localized egocentric video demonstrations to augment VLMs in two ways -- through understanding spatial task-affordances, i.e. where an agent must be for the task to physically take place, and the localization of that task relative to the egocentric viewer. We show our approach outperforms the baseline of using a VLM to map similarity of a task's description over a set of location-tagged images. Our approach has less error both on predicting where a task may take place and on predicting what tasks are likely to happen at the current location. The resulting representation will enable robots to use egocentric sensing to navigate to, or around, physical regions of interest for novel tasks specified in natural language.

Zhixuan Yu, Ziqian Bai, Abhimitra Meka et al.

Traditional methods for constructing high-quality, personalized head avatars from monocular videos demand extensive face captures and training time, posing a significant challenge for scalability. This paper introduces a novel approach to create high quality head avatar utilizing only a single or a few images per user. We learn a generative model for 3D animatable photo-realistic head avatar from a multi-view dataset of expressions from 2407 subjects, and leverage it as a prior for creating personalized avatar from few-shot images. Different from previous 3D-aware face generative models, our prior is built with a 3DMM-anchored neural radiance field backbone, which we show to be more effective for avatar creation through auto-decoding based on few-shot inputs. We also handle unstable 3DMM fitting by jointly optimizing the 3DMM fitting and camera calibration that leads to better few-shot adaptation. Our method demonstrates compelling results and outperforms existing state-of-the-art methods for few-shot avatar adaptation, paving the way for more efficient and personalized avatar creation.

Zachary Chavis, Stephen J. Guy, Hyun Soo Park

In this paper, we address the challenge of understanding human activities from an egocentric perspective. Traditional activity recognition techniques face unique challenges in egocentric videos due to the highly dynamic nature of the head during many activities. We propose a framework that seeks to address these challenges in a way that is independent of network architecture by restricting the ego-video input to a stabilized, hand-focused video. We demonstrate that this straightforward video transformation alone outperforms existing egocentric video baselines on the Ego-Exo4D Fine-Grained Keystep Recognition benchmark without requiring any alteration of the underlying model infrastructure.

Zhijian Yang, Xiaoran Fan, Volkan Isler et al.

Reconstructing the 3D pose of a person in metric scale from a single view image is a geometrically ill-posed problem. For example, we can not measure the exact distance of a person to the camera from a single view image without additional scene assumptions (e.g., known height). Existing learning based approaches circumvent this issue by reconstructing the 3D pose up to scale. However, there are many applications such as virtual telepresence, robotics, and augmented reality that require metric scale reconstruction. In this paper, we show that audio signals recorded along with an image, provide complementary information to reconstruct the metric 3D pose of the person. The key insight is that as the audio signals traverse across the 3D space, their interactions with the body provide metric information about the body's pose. Based on this insight, we introduce a time-invariant transfer function called pose kernel -- the impulse response of audio signals induced by the body pose. The main properties of the pose kernel are that (1) its envelope highly correlates with 3D pose, (2) the time response corresponds to arrival time, indicating the metric distance to the microphone, and (3) it is invariant to changes in the scene geometry configurations. Therefore, it is readily generalizable to unseen scenes. We design a multi-stage 3D CNN that fuses audio and visual signals and learns to reconstruct 3D pose in a metric scale. We show that our multi-modal method produces accurate metric reconstruction in real world scenes, which is not possible with state-of-the-art lifting approaches including parametric mesh regression and depth regression.

Kristen Grauman, Andrew Westbury, Eugene Byrne et al.

We introduce Ego4D, a massive-scale egocentric video dataset and benchmark suite. It offers 3,670 hours of daily-life activity video spanning hundreds of scenarios (household, outdoor, workplace, leisure, etc.) captured by 931 unique camera wearers from 74 worldwide locations and 9 different countries. The approach to collection is designed to uphold rigorous privacy and ethics standards with consenting participants and robust de-identification procedures where relevant. Ego4D dramatically expands the volume of diverse egocentric video footage publicly available to the research community. Portions of the video are accompanied by audio, 3D meshes of the environment, eye gaze, stereo, and/or synchronized videos from multiple egocentric cameras at the same event. Furthermore, we present a host of new benchmark challenges centered around understanding the first-person visual experience in the past (querying an episodic memory), present (analyzing hand-object manipulation, audio-visual conversation, and social interactions), and future (forecasting activities). By publicly sharing this massive annotated dataset and benchmark suite, we aim to push the frontier of first-person perception. Project page: https://ego4d-data.org/

Praneet C. Bala, Jan Zimmermann, Hyun Soo Park et al.

Recent technological developments have spurred great advances in the computerized tracking of joints and other landmarks in moving animals, including humans. Such tracking promises important advances in biology and biomedicine. Modern tracking models depend critically on labor-intensive annotated datasets of primary landmarks by non-expert humans. However, such annotation approaches can be costly and impractical for secondary landmarks, that is, ones that reflect fine-grained geometry of animals, and that are often specific to customized behavioral tasks. Due to visual and geometric ambiguity, nonexperts are often not qualified for secondary landmark annotation, which can require anatomical and zoological knowledge. These barriers significantly impede downstream behavioral studies because the learned tracking models exhibit limited generalizability. We hypothesize that there exists a shared representation between the primary and secondary landmarks because the range of motion of the secondary landmarks can be approximately spanned by that of the primary landmarks. We present a method to learn this spatial relationship of the primary and secondary landmarks in three dimensional space, which can, in turn, self-supervise the secondary landmark detector. This 3D representation learning is generic, and can therefore be applied to various multiview settings across diverse organisms, including macaques, flies, and humans.

Jae Shin Yoon, Zhixuan Yu, Jaesik Park et al.

This paper presents a new large multiview dataset called HUMBI for human body expressions with natural clothing. The goal of HUMBI is to facilitate modeling view-specific appearance and geometry of five primary body signals including gaze, face, hand, body, and garment from assorted people. 107 synchronized HD cameras are used to capture 772 distinctive subjects across gender, ethnicity, age, and style. With the multiview image streams, we reconstruct high fidelity body expressions using 3D mesh models, which allows representing view-specific appearance. We demonstrate that HUMBI is highly effective in learning and reconstructing a complete human model and is complementary to the existing datasets of human body expressions with limited views and subjects such as MPII-Gaze, Multi-PIE, Human3.6M, and Panoptic Studio datasets. Based on HUMBI, we formulate a new benchmark challenge of a pose-guided appearance rendering task that aims to substantially extend photorealism in modeling diverse human expressions in 3D, which is the key enabling factor of authentic social tele-presence. HUMBI is publicly available at http://humbi-data.net

Zhixuan Yu, Haozheng Yu, Long Sha et al.

This paper presents a new end-to-end semi-supervised framework to learn a dense keypoint detector using unlabeled multiview images. A key challenge lies in finding the exact correspondences between the dense keypoints in multiple views since the inverse of the keypoint mapping can be neither analytically derived nor differentiated. This limits applying existing multiview supervision approaches used to learn sparse keypoints that rely on the exact correspondences. To address this challenge, we derive a new probabilistic epipolar constraint that encodes the two desired properties. (1) Soft correspondence: we define a matchability, which measures a likelihood of a point matching to the other image's corresponding point, thus relaxing the requirement of the exact correspondences. (2) Geometric consistency: every point in the continuous correspondence fields must satisfy the multiview consistency collectively. We formulate a probabilistic epipolar constraint using a weighted average of epipolar errors through the matchability thereby generalizing the point-to-point geometric error to the field-to-field geometric error. This generalization facilitates learning a geometrically coherent dense keypoint detection model by utilizing a large number of unlabeled multiview images. Additionally, to prevent degenerative cases, we employ a distillation-based regularization by using a pretrained model. Finally, we design a new neural network architecture, made of twin networks, that effectively minimizes the probabilistic epipolar errors of all possible correspondences between two view images by building affinity matrices. Our method shows superior performance compared to existing methods, including non-differentiable bootstrapping in terms of keypoint accuracy, multiview consistency, and 3D reconstruction accuracy.

Yasamin Jafarian, Hyun Soo Park

A key challenge of learning a visual representation for the 3D high fidelity geometry of dressed humans lies in the limited availability of the ground truth data (e.g., 3D scanned models), which results in the performance degradation of 3D human reconstruction when applying to real-world imagery. We address this challenge by leveraging a new data resource: a number of social media dance videos that span diverse appearance, clothing styles, performances, and identities. Each video depicts dynamic movements of the body and clothes of a single person while lacking the 3D ground truth geometry. To learn a visual representation from these videos, we present a new self-supervised learning method to use the local transformation that warps the predicted local geometry of the person from an image to that of another image at a different time instant. This allows self-supervision by enforcing a temporal coherence over the predictions. In addition, we jointly learn the depths along with the surface normals that are highly responsive to local texture, wrinkle, and shade by maximizing their geometric consistency. Our method is end-to-end trainable, resulting in high fidelity depth estimation that predicts fine geometry faithful to the input real image. We further provide a theoretical bound of self-supervised learning via an uncertainty analysis that characterizes the performance of the self-supervised learning without training. We demonstrate that our method outperforms the state-of-the-art human depth estimation and human shape recovery approaches on both real and rendered images.

Jae Shin Yoon, Kihwan Kim, Jan Kautz et al.

In this paper, we present a method of clothes retargeting; generating the potential poses and deformations of a given 3D clothing template model to fit onto a person in a single RGB image. The problem is fundamentally ill-posed as attaining the ground truth data is impossible, i.e., images of people wearing the different 3D clothing template model at exact same pose. We address this challenge by utilizing large-scale synthetic data generated from physical simulation, allowing us to map 2D dense body pose to 3D clothing deformation. With the simulated data, we propose a semi-supervised learning framework that validates the physical plausibility of the 3D deformation by matching with the prescribed body-to-cloth contact points and clothing silhouette to fit onto the unlabeled real images. A new neural clothes retargeting network (CRNet) is designed to integrate the semi-supervised retargeting task in an end-to-end fashion. In our evaluation, we show that our method can predict the realistic 3D pose and deformation field needed for retargeting clothes models in real-world examples.

Jae Shin Yoon, Lingjie Liu, Vladislav Golyanik et al.

We present a new pose transfer method for synthesizing a human animation from a single image of a person controlled by a sequence of body poses. Existing pose transfer methods exhibit significant visual artifacts when applying to a novel scene, resulting in temporal inconsistency and failures in preserving the identity and textures of the person. To address these limitations, we design a compositional neural network that predicts the silhouette, garment labels, and textures. Each modular network is explicitly dedicated to a subtask that can be learned from the synthetic data. At the inference time, we utilize the trained network to produce a unified representation of appearance and its labels in UV coordinates, which remains constant across poses. The unified representation provides an incomplete yet strong guidance to generating the appearance in response to the pose change. We use the trained network to complete the appearance and render it with the background. With these strategies, we are able to synthesize human animations that can preserve the identity and appearance of the person in a temporally coherent way without any fine-tuning of the network on the testing scene. Experiments show that our method outperforms the state-of-the-arts in terms of synthesis quality, temporal coherence, and generalization ability.

Tien Do, Khiem Vuong, Stergios I. Roumeliotis et al.

In this paper, we present a spatial rectifier to estimate surface normals of tilted images. Tilted images are of particular interest as more visual data are captured by arbitrarily oriented sensors such as body-/robot-mounted cameras. Existing approaches exhibit bounded performance on predicting surface normals because they were trained using gravity-aligned images. Our two main hypotheses are: (1) visual scene layout is indicative of the gravity direction; and (2) not all surfaces are equally represented by a learned estimator due to the structured distribution of the training data, thus, there exists a transformation for each tilted image that is more responsive to the learned estimator than others. We design a spatial rectifier that is learned to transform the surface normal distribution of a tilted image to the rectified one that matches the gravity-aligned training data distribution. Along with the spatial rectifier, we propose a novel truncated angular loss that offers a stronger gradient at smaller angular errors and robustness to outliers. The resulting estimator outperforms the state-of-the-art methods including data augmentation baselines not only on ScanNet and NYUv2 but also on a new dataset called Tilt-RGBD that includes considerable roll and pitch camera motion.

Jae Shin Yoon, Kihwan Kim, Orazio Gallo et al.

This paper presents a new method to synthesize an image from arbitrary views and times given a collection of images of a dynamic scene. A key challenge for the novel view synthesis arises from dynamic scene reconstruction where epipolar geometry does not apply to the local motion of dynamic contents. To address this challenge, we propose to combine the depth from single view (DSV) and the depth from multi-view stereo (DMV), where DSV is complete, i.e., a depth is assigned to every pixel, yet view-variant in its scale, while DMV is view-invariant yet incomplete. Our insight is that although its scale and quality are inconsistent with other views, the depth estimation from a single view can be used to reason about the globally coherent geometry of dynamic contents. We cast this problem as learning to correct the scale of DSV, and to refine each depth with locally consistent motions between views to form a coherent depth estimation. We integrate these tasks into a depth fusion network in a self-supervised fashion. Given the fused depth maps, we synthesize a photorealistic virtual view in a specific location and time with our deep blending network that completes the scene and renders the virtual view. We evaluate our method of depth estimation and view synthesis on diverse real-world dynamic scenes and show the outstanding performance over existing methods.

Jae Shin Yoon, Takaaki Shiratori, Shoou-I Yu et al.

Improvements in data-capture and face modeling techniques have enabled us to create high-fidelity realistic face models. However, driving these realistic face models requires special input data, e.g. 3D meshes and unwrapped textures. Also, these face models expect clean input data taken under controlled lab environments, which is very different from data collected in the wild. All these constraints make it challenging to use the high-fidelity models in tracking for commodity cameras. In this paper, we propose a self-supervised domain adaptation approach to enable the animation of high-fidelity face models from a commodity camera. Our approach first circumvents the requirement for special input data by training a new network that can directly drive a face model just from a single 2D image. Then, we overcome the domain mismatch between lab and uncontrolled environments by performing self-supervised domain adaptation based on "consecutive frame texture consistency" based on the assumption that the appearance of the face is consistent over consecutive frames, avoiding the necessity of modeling the new environment such as lighting or background. Experiments show that we are able to drive a high-fidelity face model to perform complex facial motion from a cellphone camera without requiring any labeled data from the new domain.

Yuan Yao, Hyun Soo Park

This paper presents a semi-supervised learning framework for a customized semantic segmentation task using multiview image streams. A key challenge of the customized task lies in the limited accessibility of the labeled data due to the requirement of prohibitive manual annotation effort. We hypothesize that it is possible to leverage multiview image streams that are linked through the underlying 3D geometry, which can provide an additional supervisionary signal to train a segmentation model. We formulate a new cross-supervision method using a shape belief transfer---the segmentation belief in one image is used to predict that of the other image through epipolar geometry analogous to shape-from-silhouette. The shape belief transfer provides the upper and lower bounds of the segmentation for the unlabeled data where its gap approaches asymptotically to zero as the number of the labeled views increases. We integrate this theory to design a novel network that is agnostic to camera calibration, network model, and semantic category and bypasses the intermediate process of suboptimal 3D reconstruction. We validate this network by recognizing a customized semantic category per pixel from realworld visual data including non-human species and a subject of interest in social videos where attaining large-scale annotation data is infeasible.

Jayant Sharma, Zixing Wang, Alberto Speranzon et al.

We present a new method to localize a camera within a previously unseen environment perceived from an egocentric point of view. Although this is, in general, an ill-posed problem, humans can effortlessly and efficiently determine their relative location and orientation and navigate into a previously unseen environments, e.g., finding a specific item in a new grocery store. To enable such a capability, we design a new egocentric representation, which we call ECO (Egocentric COgnitive map). ECO is biologically inspired, by the cognitive map that allows human navigation, and it encodes the surrounding visual semantics with respect to both distance and orientation. ECO possesses three main properties: (1) reconfigurability: complex semantics and geometry is captured via the synthesis of atomic visual representations (e.g., image patch); (2) robustness: the visual semantics are registered in a geometrically consistent way (e.g., aligning with respect to the gravity vector, frontalizing, and rescaling to canonical depth), thus enabling us to learn meaningful atomic representations; (3) adaptability: a domain adaptation framework is designed to generalize the learned representation without manual calibration. As a proof-of-concept, we use ECO to localize a camera within real-world scenes---various grocery stores---and demonstrate performance improvements when compared to existing semantic localization approaches.

Zhixuan Yu, Jae Shin Yoon, In Kyu Lee et al.

This paper presents a new large multiview dataset called HUMBI for human body expressions with natural clothing. The goal of HUMBI is to facilitate modeling view-specific appearance and geometry of gaze, face, hand, body, and garment from assorted people. 107 synchronized HD cameras are used to capture 772 distinctive subjects across gender, ethnicity, age, and physical condition. With the multiview image streams, we reconstruct high fidelity body expressions using 3D mesh models, which allows representing view-specific appearance using their canonical atlas. We demonstrate that HUMBI is highly effective in learning and reconstructing a complete human model and is complementary to the existing datasets of human body expressions with limited views and subjects such as MPII-Gaze, Multi-PIE, Human3.6M, and Panoptic Studio datasets.

Yilun Zhang, Hyun Soo Park

This paper presents a semi-supervised learning framework to train a keypoint detector using multiview image streams given the limited labeled data (typically $<$4\%). We leverage the complementary relationship between multiview geometry and visual tracking to provide three types of supervisionary signals to utilize the unlabeled data: (1) keypoint detection in one view can be supervised by other views via the epipolar geometry; (2) a keypoint moves smoothly over time where its optical flow can be used to temporally supervise consecutive image frames to each other; (3) visible keypoint in one view is likely to be visible in the adjacent view. We integrate these three signals in a differentiable fashion to design a new end-to-end neural network composed of three pathways. This design allows us to extensively use the unlabeled data to train the keypoint detector. We show that our approach outperforms existing detectors including DeepLabCut tailored to the keypoint detection of non-human species such as monkeys, dogs, and mice.

Yuan Yao, Yasamin Jafarian, Hyun Soo Park

This paper presents MONET -- an end-to-end semi-supervised learning framework for a keypoint detector using multiview image streams. In particular, we consider general subjects such as non-human species where attaining a large scale annotated dataset is challenging. While multiview geometry can be used to self-supervise the unlabeled data, integrating the geometry into learning a keypoint detector is challenging due to representation mismatch. We address this mismatch by formulating a new differentiable representation of the epipolar constraint called epipolar divergence---a generalized distance from the epipolar lines to the corresponding keypoint distribution. Epipolar divergence characterizes when two view keypoint distributions produce zero reprojection error. We design a twin network that minimizes the epipolar divergence through stereo rectification that can significantly alleviate computational complexity and sampling aliasing in training. We demonstrate that our framework can localize customized keypoints of diverse species, e.g., humans, dogs, and monkeys.

Jae Shin Yoon, Ziwei Li, Hyun Soo Park

This paper presents a method to reconstruct dense semantic trajectory stream of human interactions in 3D from synchronized multiple videos. The interactions inherently introduce self-occlusion and illumination/appearance/shape changes, resulting in highly fragmented trajectory reconstruction with noisy and coarse semantic labels. Our conjecture is that among many views, there exists a set of views that can confidently recognize the visual semantic label of a 3D trajectory. We introduce a new representation called 3D semantic map---a probability distribution over the semantic labels per trajectory. We construct the 3D semantic map by reasoning about visibility and 2D recognition confidence based on view-pooling, i.e., finding the view that best represents the semantics of the trajectory. Using the 3D semantic map, we precisely infer all trajectory labels jointly by considering the affinity between long range trajectories via estimating their local rigid transformations. This inference quantitatively outperforms the baseline approaches in terms of predictive validity, representation robustness, and affinity effectiveness. We demonstrate that our algorithm can robustly compute the semantic labels of a large scale trajectory set involving real-world human interactions with object, scenes, and people.

Shan Su, Jianbo Shi, Hyun Soo Park

This paper studies a problem of inverse visual path planning: creating a visual scene from a first person action. Our conjecture is that the spatial arrangement of a first person visual scene is deployed to afford an action, and therefore, the action can be inversely used to synthesize a new scene such that the action is feasible. As a proof-of-concept, we focus on linking visual experiences induced by walking. A key innovation of this paper is a concept of ActionTunnel---a 3D virtual tunnel along the future trajectory encoding what the wearer will visually experience as moving into the scene. This connects two distinctive first person images through similar walking paths. Our method takes a first person image with a user defined future trajectory and outputs a new image that can afford the future motion. The image is created by combining present and future ActionTunnels in 3D where the missing pixels in adjoining area are computed by a generative adversarial network. Our work can provide a travel across different first person experiences in diverse real world scenes.

Shan Su, Jung Pyo Hong, Jianbo Shi et al.

This paper presents a method to predict the future movements (location and gaze direction) of basketball players as a whole from their first person videos. The predicted behaviors reflect an individual physical space that affords to take the next actions while conforming to social behaviors by engaging to joint attention. Our key innovation is to use the 3D reconstruction of multiple first person cameras to automatically annotate each other's the visual semantics of social configurations. We leverage two learning signals uniquely embedded in first person videos. Individually, a first person video records the visual semantics of a spatial and social layout around a person that allows associating with past similar situations. Collectively, first person videos follow joint attention that can link the individuals to a group. We learn the egocentric visual semantics of group movements using a Siamese neural network to retrieve future trajectories. We consolidate the retrieved trajectories from all players by maximizing a measure of social compatibility---the gaze alignment towards joint attention predicted by their social formation, where the dynamics of joint attention is learned by a long-term recurrent convolutional network. This allows us to characterize which social configuration is more plausible and predict future group trajectories.

Gedas Bertasius, Hyun Soo Park, Stella X. Yu et al.

This paper presents a method to assess a basketball player's performance from his/her first-person video. A key challenge lies in the fact that the evaluation metric is highly subjective and specific to a particular evaluator. We leverage the first-person camera to address this challenge. The spatiotemporal visual semantics provided by a first-person view allows us to reason about the camera wearer's actions while he/she is participating in an unscripted basketball game. Our method takes a player's first-person video and provides a player's performance measure that is specific to an evaluator's preference. To achieve this goal, we first use a convolutional LSTM network to detect atomic basketball events from first-person videos. Our network's ability to zoom-in to the salient regions addresses the issue of a severe camera wearer's head movement in first-person videos. The detected atomic events are then passed through the Gaussian mixtures to construct a highly non-linear visual spatiotemporal basketball assessment feature. Finally, we use this feature to learn a basketball assessment model from pairs of labeled first-person basketball videos, for which a basketball expert indicates, which of the two players is better. We demonstrate that despite not knowing the basketball evaluator's criterion, our model learns to accurately assess the players in real-world games. Furthermore, our model can also discover basketball events that contribute positively and negatively to a player's performance.

Gedas Bertasius, Hyun Soo Park, Stella X. Yu et al.

A first-person camera, placed at a person's head, captures, which objects are important to the camera wearer. Most prior methods for this task learn to detect such important objects from the manually labeled first-person data in a supervised fashion. However, important objects are strongly related to the camera wearer's internal state such as his intentions and attention, and thus, only the person wearing the camera can provide the importance labels. Such a constraint makes the annotation process costly and limited in scalability. In this work, we show that we can detect important objects in first-person images without the supervision by the camera wearer or even third-person labelers. We formulate an important detection problem as an interplay between the 1) segmentation and 2) recognition agents. The segmentation agent first proposes a possible important object segmentation mask for each image, and then feeds it to the recognition agent, which learns to predict an important object mask using visual semantics and spatial features. We implement such an interplay between both agents via an alternating cross-pathway supervision scheme inside our proposed Visual-Spatial Network (VSN). Our VSN consists of spatial ("where") and visual ("what") pathways, one of which learns common visual semantics while the other focuses on the spatial location cues. Our unsupervised learning is accomplished via a cross-pathway supervision, where one pathway feeds its predictions to a segmentation agent, which proposes a candidate important object segmentation mask that is then used by the other pathway as a supervisory signal. We show our method's success on two different important object datasets, where our method achieves similar or better results as the supervised methods.

Gedas Bertasius, Hyun Soo Park, Stella X. Yu et al.

Unlike traditional third-person cameras mounted on robots, a first-person camera, captures a person's visual sensorimotor object interactions from up close. In this paper, we study the tight interplay between our momentary visual attention and motor action with objects from a first-person camera. We propose a concept of action-objects---the objects that capture person's conscious visual (watching a TV) or tactile (taking a cup) interactions. Action-objects may be task-dependent but since many tasks share common person-object spatial configurations, action-objects exhibit a characteristic 3D spatial distance and orientation with respect to the person. We design a predictive model that detects action-objects using EgoNet, a joint two-stream network that holistically integrates visual appearance (RGB) and 3D spatial layout (depth and height) cues to predict per-pixel likelihood of action-objects. Our network also incorporates a first-person coordinate embedding, which is designed to learn a spatial distribution of the action-objects in the first-person data. We demonstrate EgoNet's predictive power, by showing that it consistently outperforms previous baseline approaches. Furthermore, EgoNet also exhibits a strong generalization ability, i.e., it predicts semantically meaningful objects in novel first-person datasets. Our method's ability to effectively detect action-objects could be used to improve robots' understanding of human-object interactions.

Gedas Bertasius, Hyun Soo Park, Jianbo Shi

On a minute-to-minute basis people undergo numerous fluid interactions with objects that barely register on a conscious level. Recent neuroscientific research demonstrates that humans have a fixed size prior for salient objects. This suggests that a salient object in 3D undergoes a consistent transformation such that people's visual system perceives it with an approximately fixed size. This finding indicates that there exists a consistent egocentric object prior that can be characterized by shape, size, depth, and location in the first person view. In this paper, we develop an EgoObject Representation, which encodes these characteristics by incorporating shape, location, size and depth features from an egocentric RGBD image. We empirically show that this representation can accurately characterize the egocentric object prior by testing it on an egocentric RGBD dataset for three tasks: the 3D saliency detection, future saliency prediction, and interaction classification. This representation is evaluated on our new Egocentric RGBD Saliency dataset that includes various activities such as cooking, dining, and shopping. By using our EgoObject representation, we outperform previously proposed models for saliency detection (relative 30% improvement for 3D saliency detection task) on our dataset. Additionally, we demonstrate that this representation allows us to predict future salient objects based on the gaze cue and classify people's interactions with objects.

Hyun Soo Park, Yedong Niu, Jianbo Shi

This paper presents a method for future localization: to predict a set of plausible trajectories of ego-motion given a depth image. We predict paths avoiding obstacles, between objects, even paths turning around a corner into space behind objects. As a byproduct of the predicted trajectories of ego-motion, we discover in the image the empty space occluded by foreground objects. We use no image based features such as semantic labeling/segmentation or object detection/recognition for this algorithm. Inspired by proxemics, we represent the space around a person using an EgoSpace map, akin to an illustrated tourist map, that measures a likelihood of occlusion at the egocentric coordinate system. A future trajectory of ego-motion is modeled by a linear combination of compact trajectory bases allowing us to constrain the predicted trajectory. We learn the relationship between the EgoSpace map and trajectory from the EgoMotion dataset providing in-situ measurements of the future trajectory. A cost function that takes into account partial occlusion due to foreground objects is minimized to predict a trajectory. This cost function generates a trajectory that passes through the occluded space, which allows us to discover the empty space behind the foreground objects. We quantitatively evaluate our method to show predictive validity and apply to various real world scenes including walking, shopping, and social interactions.