Andrew I. Comport

Cheng-You Lu, Peisen Zhou, Angela Xing et al. · stanford

Advances in neural fields are enabling high-fidelity capture of the shape and appearance of dynamic 3D scenes. However, their capabilities lag behind those offered by conventional representations such as 2D videos because of algorithmic challenges and the lack of large-scale multi-view real-world datasets. We address the dataset limitation with DiVa-360, a real-world 360 dynamic visual dataset that contains synchronized high-resolution and long-duration multi-view video sequences of table-scale scenes captured using a customized low-cost system with 53 cameras. It contains 21 object-centric sequences categorized by different motion types, 25 intricate hand-object interaction sequences, and 8 long-duration sequences for a total of 17.4 M image frames. In addition, we provide foreground-background segmentation masks, synchronized audio, and text descriptions. We benchmark the state-of-the-art dynamic neural field methods on DiVa-360 and provide insights about existing methods and future challenges on long-duration neural field capture.

Arnab Dey, Yassine Ahmine, Andrew I. Comport

Neural scene representations, such as Neural Radiance Fields (NeRF), are based on training a multilayer perceptron (MLP) using a set of color images with known poses. An increasing number of devices now produce RGB-D(color + depth) information, which has been shown to be very important for a wide range of tasks. Therefore, the aim of this paper is to investigate what improvements can be made to these promising implicit representations by incorporating depth information with the color images. In particular, the recently proposed Mip-NeRF approach, which uses conical frustums instead of rays for volume rendering, allows one to account for the varying area of a pixel with distance from the camera center. The proposed method additionally models depth uncertainty. This allows to address major limitations of NeRF-based approaches including improving the accuracy of geometry, reduced artifacts, faster training time, and shortened prediction time. Experiments are performed on well-known benchmark scenes, and comparisons show improved accuracy in scene geometry and photometric reconstruction, while reducing the training time by 3 - 5 times.

Houssem Boulahbal, Adrian Voicila, Andrew Comport

This paper proposes a self-supervised monocular image-to-depth prediction framework that is trained with an end-to-end photometric loss that handles not only 6-DOF camera motion but also 6-DOF moving object instances. Self-supervision is performed by warping the images across a video sequence using depth and scene motion including object instances. One novelty of the proposed method is the use of the multi-head attention of the transformer network that matches moving objects across time and models their interaction and dynamics. This enables accurate and robust pose estimation for each object instance. Most image-to-depth predication frameworks make the assumption of rigid scenes, which largely degrades their performance with respect to dynamic objects. Only a few SOTA papers have accounted for dynamic objects. The proposed method is shown to outperform these methods on standard benchmarks and the impact of the dynamic motion on these benchmarks is exposed. Furthermore, the proposed image-to-depth prediction framework is also shown to be competitive with SOTA video-to-depth prediction frameworks.

Houssem Boulahbal, Adrian Voicila, Andrew Comport

This paper addresses the problem of end-to-end self-supervised forecasting of depth and ego motion. Given a sequence of raw images, the aim is to forecast both the geometry and ego-motion using a self supervised photometric loss. The architecture is designed using both convolution and transformer modules. This leverages the benefits of both modules: Inductive bias of CNN, and the multi-head attention of transformers, thus enabling a rich spatio-temporal representation that enables accurate depth forecasting. Prior work attempts to solve this problem using multi-modal input/output with supervised ground-truth data which is not practical since a large annotated dataset is required. Alternatively to prior methods, this paper forecasts depth and ego motion using only self-supervised raw images as input. The approach performs significantly well on the KITTI dataset benchmark with several performance criteria being even comparable to prior non-forecasting self-supervised monocular depth inference methods.

Yassine Ahmine, Arnab Dey, Andrew I. Comport

Recently neural scene representations have provided very impressive results for representing 3D scenes visually, however, their study and progress have mainly been limited to visualization of virtual models in computer graphics or scene reconstruction in computer vision without explicitly accounting for sensor and pose uncertainty. Using this novel scene representation in robotics applications, however, would require accounting for this uncertainty in the neural map. The aim of this paper is therefore to propose a novel method for training {\em probabilistic neural scene representations} with uncertain training data that could enable the inclusion of these representations in robotics applications. Acquiring images using cameras or depth sensors contains inherent uncertainty, and furthermore, the camera poses used for learning a 3D model are also imperfect. If these measurements are used for training without accounting for their uncertainty, then the resulting models are non-optimal, and the resulting scene representations are likely to contain artifacts such as blur and un-even geometry. In this work, the problem of uncertainty integration to the learning process is investigated by focusing on training with uncertain information in a probabilistic manner. The proposed method involves explicitly augmenting the training likelihood with an uncertainty term such that the learnt probability distribution of the network is minimized with respect to the training uncertainty. It will be shown that this leads to more accurate image rendering quality, in addition to more precise and consistent geometry. Validation has been carried out on both synthetic and real datasets showing that the proposed approach outperforms state-of-the-art methods. The results show notably that the proposed method is capable of rendering novel high-quality views even when the training data is limited.

Houssem Boulahbal, Adrian Voicila, Andrew Comport

In this paper, a self-supervised model that simultaneously predicts a sequence of future frames from video-input with a novel spatial-temporal attention (ST) network is proposed. The ST transformer network allows constraining both temporal consistency across future frames whilst constraining consistency across spatial objects in the image at different scales. This was not the case in prior works for depth prediction, which focused on predicting a single frame as output. The proposed model leverages prior scene knowledge such as object shape and texture similar to single-image depth inference methods, whilst also constraining the motion and geometry from a sequence of input images. Apart from the transformer architecture, one of the main contributions with respect to prior works lies in the objective function that enforces spatio-temporal consistency across a sequence of output frames rather than a single output frame. As will be shown, this results in more accurate and robust depth sequence forecasting. The model achieves highly accurate depth forecasting results that outperform existing baselines on the KITTI benchmark. Extensive ablation studies were performed to assess the effectiveness of the proposed techniques. One remarkable result of the proposed model is that it is implicitly capable of forecasting the motion of objects in the scene, rather than requiring complex models involving multi-object detection, segmentation and tracking.

Arnab Dey, Andrew I. Comport

Learning a 3D representation of a scene has been a challenging problem for decades in computer vision. Recent advances in implicit neural representation from images using neural radiance fields(NeRF) have shown promising results. Some of the limitations of previous NeRF based methods include longer training time, and inaccurate underlying geometry. The proposed method takes advantage of RGB-D data to reduce training time by leveraging depth sensing to improve local sampling. This paper proposes a depth-guided local sampling strategy and a smaller neural network architecture to achieve faster training time without compromising quality.

Antoine Schnepf, Karim Kassab, Jean-Yves Franceschi et al.

While pre-trained image autoencoders are increasingly utilized in computer vision, the application of inverse graphics in 2D latent spaces has been under-explored. Yet, besides reducing the training and rendering complexity, applying inverse graphics in the latent space enables a valuable interoperability with other latent-based 2D methods. The major challenge is that inverse graphics cannot be directly applied to such image latent spaces because they lack an underlying 3D geometry. In this paper, we propose an Inverse Graphics Autoencoder (IG-AE) that specifically addresses this issue. To this end, we regularize an image autoencoder with 3D-geometry by aligning its latent space with jointly trained latent 3D scenes. We utilize the trained IG-AE to bring NeRFs to the latent space with a latent NeRF training pipeline, which we implement in an open-source extension of the Nerfstudio framework, thereby unlocking latent scene learning for its supported methods. We experimentally confirm that Latent NeRFs trained with IG-AE present an improved quality compared to a standard autoencoder, all while exhibiting training and rendering accelerations with respect to NeRFs trained in the image space. Our project page can be found at https://ig-ae.github.io .

Karim Kassab, Antoine Schnepf, Jean-Yves Franceschi et al.

To learn large sets of scenes, Tri-Planes are commonly employed for their planar structure that enables an interoperability with image models, and thus diverse 3D applications. However, this advantage comes at the cost of resource efficiency, as Tri-Planes are not the most computationally efficient option. In this paper, we introduce Fused-Planes, a new planar architecture that improves Tri-Planes resource-efficiency in the framework of learning large sets of scenes, which we call "multi-scene inverse graphics". To learn a large set of scenes, our method divides it into two subsets and operates as follows: (i) we train the first subset of scenes jointly with a compression model, (ii) we use that compression model to learn the remaining scenes. This compression model consists of a 3D-aware latent space in which Fused-Planes are learned, enabling a reduced rendering resolution, and shared structures across scenes that reduce scene representation complexity. Fused-Planes present competitive resource costs in multi-scene inverse graphics, while preserving Tri-Planes rendering quality, and maintaining their widely favored planar structure. Our codebase is publicly available as open-source. Our project page can be found at https://fused-planes.github.io .

Arnab Dey, Cheng-You Lu, Andrew I. Comport et al.

Recent advancements in radiance field rendering show promising results in 3D scene representation, where Gaussian splatting-based techniques emerge as state-of-the-art due to their quality and efficiency. Gaussian splatting is widely used for various applications, including 3D human representation. However, previous 3D Gaussian splatting methods either use parametric body models as additional information or fail to provide any underlying structure, like human biomechanical features, which are essential for different applications. In this paper, we present a novel approach called HFGaussian that can estimate novel views and human features, such as the 3D skeleton, 3D key points, and dense pose, from sparse input images in real time at 25 FPS. The proposed method leverages generalizable Gaussian splatting technique to represent the human subject and its associated features, enabling efficient and generalizable reconstruction. By incorporating a pose regression network and the feature splatting technique with Gaussian splatting, HFGaussian demonstrates improved capabilities over existing 3D human methods, showcasing the potential of 3D human representations with integrated biomechanics. We thoroughly evaluate our HFGaussian method against the latest state-of-the-art techniques in human Gaussian splatting and pose estimation, demonstrating its real-time, state-of-the-art performance.

Arnab Dey, Di Yang, Rohith Agaram et al.

Recent advances in Neural Radiance Fields (NeRF) have demonstrated promising results in 3D scene representations, including 3D human representations. However, these representations often lack crucial information on the underlying human pose and structure, which is crucial for AR/VR applications and games. In this paper, we introduce a novel approach, termed GHNeRF, designed to address these limitations by learning 2D/3D joint locations of human subjects with NeRF representation. GHNeRF uses a pre-trained 2D encoder streamlined to extract essential human features from 2D images, which are then incorporated into the NeRF framework in order to encode human biomechanic features. This allows our network to simultaneously learn biomechanic features, such as joint locations, along with human geometry and texture. To assess the effectiveness of our method, we conduct a comprehensive comparison with state-of-the-art human NeRF techniques and joint estimation algorithms. Our results show that GHNeRF can achieve state-of-the-art results in near real-time.

Antoine Schnepf, Karim Kassab, Jean-Yves Franceschi et al.

We present a method enabling the scaling of NeRFs to learn a large number of semantically-similar scenes. We combine two techniques to improve the required training time and memory cost per scene. First, we learn a 3D-aware latent space in which we train Tri-Plane scene representations, hence reducing the resolution at which scenes are learned. Moreover, we present a way to share common information across scenes, hence allowing for a reduction of model complexity to learn a particular scene. Our method reduces effective per-scene memory costs by 44% and per-scene time costs by 86% when training 1000 scenes. Our project page can be found at https://3da-ae.github.io .

Houssem eddine Boulahbal, Adrian Voicila, Andrew Comport

This paper proposes two important contributions for conditional Generative Adversarial Networks (cGANs) to improve the wide variety of applications that exploit this architecture. The first main contribution is an analysis of cGANs to show that they are not explicitly conditional. In particular, it will be shown that the discriminator and subsequently the cGAN does not automatically learn the conditionality between inputs. The second contribution is a new method, called a contrario cGAN, that explicitly models conditionality for both parts of the adversarial architecture via a novel a contrario loss that involves training the discriminator to learn unconditional (adverse) examples. This leads to a novel type of data augmentation approach for GANs (a contrario learning) which allows to restrict the search space of the generator to conditional outputs using adverse examples. Extensive experimentation is carried out to evaluate the conditionality of the discriminator by proposing a probability distribution analysis. Comparisons with the cGAN architecture for different applications show significant improvements in performance on well known datasets including, semantic image synthesis, image segmentation, monocular depth prediction and "single label"-to-image using different metrics including Fréchet Inception Distance (FID), mean Intersection over Union (mIoU), Root Mean Square Error log (RMSE log) and Number of statistically-Different Bins (NDB).