As 3D point clouds become the representation of choice for multiple vision and graphics applications, ability to synthesize or reconstruct high-resolution, high-fidelity becomes crucial. Despite recent success deep learning models in discriminative tasks clouds, generating remains challenging. This paper proposes a principled probabilistic framework generate by modeling them as distribution distributions. Specifically, we learn two-level hierarchy distributions where first level is shapes...

Evaluation metrics for image captioning face two challenges. Firstly, commonly used such as CIDEr, METEOR, ROUGE and BLEU often do not correlate well with human judgments. Secondly, each metric has known blind spots to pathological caption constructions, rule-based lack provisions repair once identified. For example, the newly proposed SPICE correlates judgments, but fails capture syntactic structure of a sentence. To address these challenges, we propose novel learning based discriminative...

No abstract available.

In this paper, we present a self-calibrating framework that jointly optimizes camera parameters, lens distortion and 3D Gaussian representations, enabling accurate efficient scene reconstruction. particular, our technique enables high-quality reconstruction from Large field-of-view (FOV) imagery taken with wide-angle lenses, allowing the to be modeled smaller number of images. Our approach introduces novel method for modeling complex distortions using hybrid network combines invertible...

Low-precision training reduces computational cost and produces efficient models. Recent research in developing new low-precision algorithms often relies on simulation to empirically evaluate the statistical effects of quantization while avoiding substantial overhead building specific hardware. To support this empirical research, we introduce QPyTorch, a arithmetic framework. Built natively PyTorch, QPyTorch provides convenient interface that minimizes efforts needed reliably convert existing...

Low precision operations can provide scalability, memory savings, portability, and energy efficiency. This paper proposes SWALP, an approach to low training that averages low-precision SGD iterates with a modified learning rate schedule. SWALP is easy implement match the performance of full-precision even all numbers quantized down 8 bits, including gradient accumulators. Additionally, we show converges arbitrarily close optimal solution for quadratic objectives, noise ball asymptotically...

As 3D point clouds become the representation of choice for multiple vision and graphics applications, ability to synthesize or reconstruct high-resolution, high-fidelity becomes crucial. Despite recent success deep learning models in discriminative tasks clouds, generating remains challenging. This paper proposes a principled probabilistic framework generate by modeling them as distribution distributions. Specifically, we learn two-level hierarchy distributions where first level is shapes...

We introduce a general framework for solving partial differential equations (PDEs) using generative diffusion models. In particular, we focus on the scenarios where do not have full knowledge of scene necessary to apply classical solvers. Most existing forward or inverse PDE approaches perform poorly when observations data underlying coefficients are incomplete, which is common assumption real-world measurements. this work, propose DiffusionPDE that can simultaneously fill in missing...

Neural fields have emerged as a new paradigm for representing signals, thanks to their ability do it compactly while being easy optimize. In most applications, however, neural are treated like black boxes, which precludes many signal manipulation tasks. this paper, we propose class of called polynomial (PNFs). The key advantage PNF is that can represent composition number manipulable and interpretable components without losing the merits representation. We develop general theoretical...

State-of-the-art deep reading comprehension models are dominated by recurrent neural nets. Their sequential nature is a natural fit for language, but it also precludes parallelization within an instances and often becomes the bottleneck deploying such to latency critical scenarios. This particularly problematic longer texts. Here we present convolutional architecture as alternative these architectures. Using simple dilated units in place of ones, achieve results comparable state art on two...

3D Gaussians, as a low-level scene representation, typically involve thousands to millions of Gaussians. This makes it difficult control the in ways that reflect underlying dynamic structure, where number independent entities is much smaller. In particular, can be challenging animate and move objects scene, which requires coordination among many To address this issue, we develop mutual information shaping technique enforces movement resonance between correlated Gaussians motion network. Such...

Evaluation metrics for image captioning face two challenges. Firstly, commonly used such as CIDEr, METEOR, ROUGE and BLEU often do not correlate well with human judgments. Secondly, each metric has known blind spots to pathological caption constructions, rule-based lack provisions repair once identified. For example, the newly proposed SPICE correlates judgments, but fails capture syntactic structure of a sentence. To address these challenges, we propose novel learning based discriminative...

Low-precision training reduces computational cost and produces efficient models. Recent research in developing new low-precision algorithms often relies on simulation to empirically evaluate the statistical effects of quantization while avoiding substantial overhead building specific hardware. To support this empirical research, we introduce QPyTorch, a arithmetic framework. Built natively PyTorch, QPyTorch provides convenient interface that minimizes efforts needed reliably convert existing...

In this work, we propose a novel technique to generate shapes from point cloud data. A can be viewed as samples distribution of 3D points whose density is concentrated near the surface shape. Point generation thus amounts moving randomly sampled high-density areas. We clouds by performing stochastic gradient ascent on an unnormalized probability density, thereby toward high-likelihood regions. Our model directly predicts log field and trained with simple objective adapted score-based...

Neural radiance fields (NeRF) rely on volume rendering to synthesize novel views. Volume requires evaluating an integral along each ray, which is numerically approximated with a finite sum that corresponds the exact ray under piecewise constant density. As consequence, rendered result unstable w.r.t. choice of samples phenomenon we dub quadrature instability. We propose mathematically principled solution by reformulating sample-based equation so it linear This simultaneously resolves...

This paper presents a method that uses neural networks as caching mechanism to reduce the variance of Monte Carlo Partial Differential Equation solvers, such Walk-on-Spheres algorithm [Sawhney and Crane 2020]. While these PDE solvers have merits being unbiased discretization-free, their high often hinders real-time applications. On other hand, can approximate solution, evaluating at inference time be very fast. However, neural-network-based solutions may suffer from convergence difficulties...