Recent advances in neural rendering have improved both training and rendering times by orders of magnitude. While these methods demonstrate state-of-the-art quality and speed, they are designed for photogrammetry of static scenes and do not generalize well to freely moving humans in the environment. In this work, we introduce Human Gaussian Splats (HUGS) that represents an animatable human together with the scene using 3D Gaussian Splatting (3DGS). Our method takes only a monocular video with a small number of (50–100) frames, and it automatically learns to disentangle the static scene and a fully animatable human avatar within 30 minutes. We utilize the SMPL body model to initialize the human Gaussians. To capture details that are not modeled by SMPL (e.g., cloth, hairs), we allow the 3D Gaussians to deviate from the human body model. Utilizing 3D Gaussians for animated humans brings new challenges, including the artifacts created when articulating the Gaussians. We propose to jointly optimize the linear blend skinning weights to coordinate the movements of individual Gaussians during animation. Our approach enables novel-pose synthesis of human and novel view synthesis of both the human and the scene. We achieve state-of-the-art rendering quality with a rendering speed of 60 FPS while being ~100x faster to train over previous work. Our code will be announced here: https://github.com/apple/ml-hugs.

We consider the task of animating 3D facial geometry from speech signal. Existing works are primarily deterministic, focusing on learning a one-to-one mapping from speech signal to 3D face meshes on small datasets with limited speakers. While these models can achieve high-quality lip articulation for speakers in the training set, they are unable to capture the full and diverse distribution of 3D facial motions that accompany speech in the real world. Importantly, the relationship between speech and facial motion is one-to-many, containing both inter-speaker and intra-speaker variations and necessitating a probabilistic approach. In this paper, we identify and address key challenges that have so far limited the development of probabilistic models: lack of datasets and metrics that are suitable for training and evaluating them, as well as the difficulty of designing a model that generates diverse results while remaining faithful to a strong conditioning signal as speech. We first propose large-scale benchmark datasets and metrics suitable for probabilistic modeling. Then, we demonstrate a probabilistic model that achieves both diversity and fidelity to speech, outperforming other methods across the proposed benchmarks. Finally, we showcase useful applications of probabilistic models trained on these large-scale datasets: we can generate diverse speech-driven 3D facial motion that matches unseen speaker styles extracted from reference clips; and our synthetic meshes can be used to improve the performance of downstream audio-visual models.

While Automatic Speech Recognition (ASR) systems are widely used in many real-world applications, they often do not generalize well to new domains and need to be fine-tuned on data from these domains. However, target-domain data usually are not readily available in many scenarios. In this paper, we propose a new strategy for adapting ASR models to new target domains without any text or speech from those domains. To accomplish this, we propose a novel data synthesis pipeline that uses a Large Language Model (LLM) to generate a target domain text corpus, and a state-of-the-art controllable speech synthesis model to generate the corresponding speech. We propose a simple yet effective in-context instruction fine-tuning strategy to increase the effectiveness of LLM in generating text corpora for new domains. Experiments on the SLURP dataset show that the proposed method achieves an average relative word error rate improvement of 28% on unseen target domains without any performance drop in source domains.

The task of novel view synthesis aims to generate unseen perspectives of an object or scene from a limited set of input images. Nevertheless, synthesizing novel views from a single image remains a significant challenge. Previous approaches tackle this problem by adopting mesh prediction, multi-plane image construction, or more advanced techniques such as neural radiance fields. Recently, a pre-trained diffusion model that is specifically designed for 2D image synthesis has demonstrated its capability in producing photorealistic novel views, if sufficiently optimized with a 3D finetuning task. Despite greatly improved fidelity and generalizability, training such a powerful diffusion model requires a vast volume of training data and model parameters, resulting in a notoriously long time and high computational costs. To tackle this issue, we propose Efficient-3DiM, a highly efficient yet effective framework to learn a single-image novel-view synthesizer. Motivated by our in-depth analysis of the diffusion model inference process, we propose several pragmatic strategies to reduce training overhead to a manageable scale, including a crafted timestep sampling strategy, a superior 3D feature extractor, and an enhanced training scheme. When combined, our framework can reduce the total training time from 10 days to less than 1 day, significantly accelerating the training process on the same computational platform (an instance with 8 Nvidia A100 GPUs). Comprehensive experiments are conducted to demonstrate the efficiency and generalizability of our proposed method.

Large language models (LLMs) are commonly trained on datasets consisting of fixed-length token sequences. These datasets are created by randomly concatenating documents of various lengths and then chunking them into sequences of a predetermined target length (concat-and-chunk). Recent attention implementations mask cross-document attention, reducing the effective length of a chunk of tokens. Additionally, training on long sequences becomes computationally prohibitive due to the quadratic cost of attention. In this study, we introduce dataset decomposition, a novel variable sequence length training technique, to tackle these challenges. We decompose a dataset into a union of buckets, each containing sequences of the same size extracted from a unique document. During training, we use variable sequence length and batch-size, sampling simultaneously from all buckets with a curriculum. In contrast to the concat-and-chunk baseline, which incurs a fixed attention cost at every step of training, our proposed method incurs a computational cost proportional to the actual document lengths at each step, resulting in significant savings in training time. We train an 8k context-length 1B model at the same cost as a 2k context-length model trained with the baseline approach. Experiments on a web-scale corpus demonstrate that our approach significantly enhances performance on standard language evaluations and long-context benchmarks, reaching target accuracy with up to 6x faster training compared to the baseline. Our method not only enables efficient pretraining on long sequences but also scales effectively with dataset size. Lastly, we shed light on a critical yet less studied aspect of training large language models: the distribution and curriculum of sequence lengths, which results in a non-negligible difference in performance.

We propose a novel method that renders point clouds as if they are surfaces. The proposed method is differentiable and requires no scene-specific optimization. This unique capability enables, out-of-the-box, surface normal estimation, rendering room-scale point clouds, inverse rendering, and ray tracing with global illumination. Unlike existing work that focuses on converting point clouds to other representations—e.g., surfaces or implicit functions—our key idea is to directly infer the intersection of a light ray with the underlying surface represented by the given point cloud. Specifically, we train a set transformer that, given a small number of local neighbor points along a light ray, provides the intersection point, the surface normal, and the material blending weights, which are used to render the outcome of this light ray. Localizing the problem into small neighborhoods enables us to train a model with only 48 meshes and apply it to unseen point clouds. Our model achieves higher estimation accuracy than state-of-the-art surface reconstruction and point-cloud rendering methods on three test sets. When applied to room-scale point clouds, without any scene-specific optimization, the model achieves competitive quality with the state-of-the-art novel-view rendering methods. Moreover, we demonstrate ability to render and manipulate Lidar-scanned point clouds such as lighting control and object insertion.

We propose a generative framework, FaceLit, capable of generating a 3D face that can be rendered at various user-defined lighting conditions and views, learned purely from 2D images in-the-wild without any manual annotation. Unlike existing works that require careful capture setup or human labor, we rely on off-the-shelf pose and illumination estimators. With these estimates, we incorporate the Phong reflectance model in the neural volume rendering framework. Our model learns to generate shape and material properties of a face such that, when rendered according to the natural statistics of pose and illumination, produces photorealistic face images with multiview 3D and illumination consistency. Our method enables photorealistic generation of faces with explicit illumination and view controls on multiple datasets - FFHQ, MetFaces and CelebA-HQ. We show state-of-the-art photorealism among 3D aware GANs on FFHQ dataset achieving an FID score of 3.5.

Adapting generic speech recognition models to specific individuals is a challenging problem due to the scarcity of personalized data. Recent works have proposed boosting the amount of training data using personalized text-to-speech synthesis. Here, we ask two fundamental questions about this strategy: when is synthetic data effective for personalization, and why is it effective in those cases? To address the first question, we adapt a state-of-the-art automatic speech recognition (ASR) model to target speakers from four benchmark datasets representative of different speaker types. We show that ASR personalization with synthetic data is effective in all cases, but particularly when (i) the target speaker is underrepresented in the global data, and (ii) the capacity of the global model is limited. To address the second question of why personalized synthetic data is effective, we use controllable speech synthesis (CSS) to generate speech with varied styles and content. Surprisingly, we find that the text content of the synthetic data, rather than style, is important for speaker adaptation. These results lead us to propose a data selection strategy for ASR personalization based on speech content.

Pooling is commonly used to improve the computation-accuracy trade-off of convolutional networks. By aggregating neighboring feature values on the image grid, pooling layers downsample feature maps while maintaining accuracy. In standard vision transformers, however, tokens are processed individually and do not necessarily lie on regular grids. Utilizing pooling methods designed for image grids (e.g., average pooling) thus can be sub-optimal for transformers, as shown by our experiments. In this paper, we propose Token Pooling to downsample token sets in vision transformers. We take a new perspective — instead of assuming tokens form a regular grid, we treat them as discrete (and irregular) samples of an implicit continuous signal. Given a target number of tokens, Token Pooling finds the set of tokens that best approximates the underlying continuous signal. We rigorously evaluate the proposed method on the standard transformer architecture (ViT/DeiT) and on the image classification problem using ImageNet-1k. Our experiments show that Token Pooling significantly improves the computation-accuracy trade-off without any further modifications to the architecture. Token Pooling enables DeiT-Ti to achieve the same top-1 accuracy while using 42% fewer computations.

In this paper, we demonstrate how a generative model can be used to build a better recognizer through the control of content and style. We are building an online handwriting recognizer from a modest amount of training samples. By training our controllable handwriting synthesizer on the same data, we can synthesize handwriting with previously underrepresented content (e.g., URLs and email addresses) and style (e.g., cursive and slanted). Moreover, we propose a framework to analyze a recognizer that is trained with a mixture of real and synthetic training data. We use the framework to optimize data synthesis and demonstrate significant improvement on handwriting recognition over a model trained on real data only. Overall, we achieve a 66% reduction in Character Error Rate.

With recent advances in speech synthesis, synthetic data is becoming a viable alternative to real data for training speech recognition models. However, machine learning with synthetic data is not trivial due to the gap between the synthetic and the real data distributions. Synthetic datasets may contain artifacts that do not exist in real data such as structured noise, content errors, or unrealistic speaking styles. Moreover, the synthesis process may introduce a bias due to uneven sampling of the data manifold. We propose two novel techniques during training to mitigate the problems due to the distribution gap: (i) a rejection sampling algorithm and (ii) using separate batch normalization statistics for the real and the synthetic samples. We show that these methods significantly improve the training of speech recognition models using synthetic data. We evaluate the proposed approach on keyword detection and Automatic Speech Recognition (ASR) tasks, and observe up to 18% and 13% relative error reduction, respectively, compared to naively using the synthetic data.

Controllable generative sequence models with the capability to extract and replicate the style of specific examples enable many applications, including narrating audiobooks in different voices, auto-completing and auto-correcting written handwriting, and generating missing training samples for downstream recognition tasks. However, under an unsupervised-style setting, typical training algorithms for controllable sequence generative models suffer from the training-inference mismatch, where the same sample is used as content and style input during training but unpaired samples are given during inference. In this paper, we tackle the training-inference mismatch encountered during unsupervised learning of controllable generative sequence models. The proposed method is simple yet effective, where we use a style transformation module to transfer target style information into an unrelated style input. This method enables training using unpaired content and style samples and thereby mitigate the training-inference mismatch. We apply style equalization to text-to-speech and text-to-handwriting synthesis on three datasets. We conduct thorough evaluation, including both quantitative and qualitative user studies. Our results show that by mitigating the training-inference mismatch with the proposed style equalization, we achieve style replication scores comparable to real data in our user studies.

Data augmentation methods usually apply the same augmentation (or a mix of them) to all the training samples. For example, to perturb data with noise, the noise is sampled from a Normal distribution with a fixed standard deviation, for all samples. We hypothesize that a hard sample with high training loss already provides strong training signal to update the model parameters and should be perturbed with mild or no augmentation. Perturbing a hard sample with a strong augmentation may also make it too hard to learn from. Furthermore, a sample with low training loss should be perturbed by a stronger augmentation to provide more robustness to a variety of conditions. To formalize these intuitions, we propose a novel method to learn a Sample-Adaptive Policy for Augmentation – SapAugment. Our policy adapts the augmentation parameters based on the training loss of the data samples. In the example of Gaussian noise, a hard sample will be perturbed with a low variance noise and an easy sample with a high variance noise. Furthermore, the proposed method combines multiple augmentation methods into a methodical policy learning framework and obviates hand-crafting augmentation parameters by trial-and-error. We apply our method on an automatic speech recognition (ASR) task, and combine existing and novel augmentations using the proposed framework. We show substantial improvement, up to 21% relative reduction in word error rate on LibriSpeech dataset, over the state-of-the-art speech augmentation method.

One of the most promising approaches for automated detection of guns and other prohibited items in aviation baggage screening is the use of 3D computed tomography (CT) scans. However, automated detection, especially with deep neural networks, faces two key challenges: the high dimensionality of individual 3D scans, and the lack of labelled training data. We address these challenges using a novel image-based detection and segmentation technique that we call the slice-and-fuse framework. Our approach relies on slicing the input 3D volumes, generating 2D predictions on each slice using 2D Convolutional Neural Networks (CNNs), and fusing them to obtain a 3D prediction. We develop two distinct detectors based on this slice-and-fuse strategy: the Retinal-SliceNet that uses a unified, single network with end-to-end training, and the U-SliceNet that uses a two-stage paradigm, first generating proposals using a voxel labeling network and, subsequently, refining the proposals by a 3D classification network. The networks are trained using a data augmentation approach that creates a very large training dataset by inserting weapons into 3D CT scans of threat-free bags. We demonstrate that the two SliceNets outperform state-of-the-art methods on a large-scale 3D baggage CT dataset for baggage classification, 3D object detection, and 3D semantic segmentation.

The human visual system uses numerous cues for depth perception, including disparity, accommodation, motion parallax and occlusion. It is incumbent upon virtual-reality displays to satisfy these cues to provide an immersive user experience. Multifocal displays, one of the classic approaches to satisfy the accommodation cue, place virtual content at multiple focal planes, each at a different depth. However, the content on focal planes close to the eye do not occlude those farther away; this deteriorates the occlusion cue as well as reduces contrast at depth discontinuities due to leakage of the defocus blur. This paper enables occlusion-aware multifocal displays using a novel ConeTilt operator that provides an additional degree of freedom --- tilting the light cone emitted at each pixel of the display panel. We show that, for scenes with relatively simple occlusion configurations, tilting the light cones provides the same effect as physical occlusion. We demonstrate that ConeTilt can be easily implemented by a phase-only spatial light modulator. Using a lab prototype, we show results that demonstrate the presence of occlusion cues and the increased contrast of the display at depth edges.

We present a virtual reality display that is capable of generating a dense collection of depth/focal planes. This is achieved by driving a focus-tunable lens to sweep a range of focal lengths at a high frequency and, subsequently, tracking the focal length precisely at microsecond time resolutions using an optical module. Precise tracking of the focal length, coupled with a high-speed display, enables our lab prototype to generate 1600 focal planes per second. This enables a novel first-of-its-kind virtual reality multifocal display that is capable of resolving the vergence-accommodation conflict endemic to today's displays.

While deep learning methods have achieved state-of-theart performance in many challenging inverse problems like image inpainting and super-resolution, they invariably involve problem-specific training of the networks. Under this approach, each inverse problem requires its own dedicated network. In scenarios where we need to solve a wide variety of problems, e.g., on a mobile camera, it is inefficient and expensive to use these problem-specific networks. On the other hand, traditional methods using analytic signal priors can be used to solve any linear inverse problem; this often comes with a performance that is worse than learning-based methods. In this work, we provide a middle ground between the two kinds of methods - we propose a general framework to train a single deep neural network that solves arbitrary linear inverse problems. We achieve this by training a network that acts as a quasi-projection operator for the set of natural images and show that any linear inverse problem involving natural images can be solved using iterative methods. We empirically show that the proposed framework demonstrates superior performance over traditional methods using wavelet sparsity prior while achieving performance comparable to specially-trained networks on tasks including compressive sensing and pixel-wise inpainting.

Random features is an approach for kernel-based inference on large datasets. In this paper, we derive performance guarantees for random features on signals, like images, that enjoy sparse representations and show that the number of random features required to achieve a desired approximation of the kernel similarity matrix can be significantly smaller for sparse signals. Based on this, we propose a scheme termed compressive random features that first obtains low-dimensional projections of a dataset and, subsequently, derives random features on the low-dimensional projections. This scheme provides significant improvements in signal dimensionality, computational time, and storage costs over traditional random features while enjoying similar theoretical guarantees for achieving inference performance. We support our claims by providing empirical results across many datasets.

We propose the propagation filter as a novel image filtering operator, with the goal of smoothing over neighboring image pixels while preserving image context like edges or textural regions. In particular, our filter does not to utilize explicit spatial kernel functions as bilateral and guided filters do. We will show that our propagation filter can be viewed as a robust estimator, which minimizes the expected difference between the filtered and desirable image outputs. We will also relate propagation filtering to belief propagation, and suggest techniques if further speedup of the filtering process is necessary. In our experiments, we apply our propagation filter to a variety of applications such as image denoising, smoothing, fusion, and high-dynamic-range (HDR) compression. We will show that improved performance over existing image filters can be achieved.