Reward finetuning has emerged as a promising approach to aligning foundation models with downstream objectives. Remarkable success has been achieved in the language domain by using reinforcement learning (RL) to maximize rewards that reflect human preference. However, in the vision domain, existing RL-based reward finetuning methods are limited by their instability in large-scale training, rendering them incapable of generalizing to complex, unseen prompts. In this paper, we propose Proximal Reward Difference Prediction (PRDP), enabling stable black-box reward finetuning for diffusion models for the first time on large-scale prompt datasets with over 100K prompts. Our key innovation is the Reward Difference Prediction (RDP) objective that has the same optimal solution as the RL objective while enjoying better training stability. Specifically, the RDP objective is a supervised regression objective that tasks the diffusion model with predicting the reward difference of generated image pairs from their denoising trajectories. We theoretically prove that the diffusion model that obtains perfect reward difference prediction is exactly the maximizer of the RL objective. We further develop an online algorithm with proximal updates to stably optimize the RDP objective. In experiments, we demonstrate that PRDP can match the reward maximization ability of well-established RL-based methods in small-scale training. Furthermore, through large-scale training on text prompts from the Human Preference Dataset v2 and the Pick-a-Pic v1 dataset, PRDP achieves superior generation quality on a diverse set of complex, unseen prompts whereas RL-based methods completely fail.

We present StreamVC, a streaming voice conversion solution that preserves the content and prosody of any source speech while matching the voice timbre from any target speech. Unlike previous approaches, StreamVC produces the resulting waveform at low latency from the input signal even on a mobile platform, making it applicable to real-time communication scenarios like calls and video conferencing, and addressing use cases such as voice anonymization in these scenarios. Our design leverages the architecture and training strategy of the SoundStream neural audio codec for lightweight high-quality speech synthesis. We demonstrate the feasibility of learning soft speech units causally, as well as the effectiveness of supplying whitened fundamental frequency information to improve pitch stability without leaking the source timbre information.

We propose a neural network model that can separate target speech sources from interfering sources at different angular regions using two microphones. The model is trained with simulated room impulse responses (RIRs) using omnidirectional microphones without needing to collect real RIRs. By relying on specific angular regions and multiple room simulations, the model utilizes consistent time difference of arrival (TDOA) cues, or what we call delay contrast, to separate target and interference sources while remaining robust in various reverberation environments. We demonstrate the model is not only generalizable to a commercially available device with a slightly different microphone geometry, but also outperforms our previous work which uses one additional microphone on the same device. The model runs in real-time on-device and is suitable for low-latency streaming applications such as telephony and video conferencing.

High quality speech capture has been widely studied for both voice communication and human computer interface reasons. To improve the capture performance, we can often find multi-microphone speech enhancement techniques deployed on various devices. Multi-microphone speech enhancement problem is often decomposed into two decoupled steps: a beamformer that provides spatial filtering and a single-channel speech enhancement model that cleans up the beamformer output. In this work, we propose a speech enhancement solution that takes both the raw microphone and beamformer outputs as the input for an ML model. We devise a simple yet effective training scheme that allows the model to learn from the cues of the beamformer by contrasting the two inputs and greatly boost its capability in spatial rejection, while conducting the general tasks of denoising and dereverberation. The proposed solution takes advantage of classical spatial filtering algorithms instead of competing with them. By design, the beamformer module then could be selected separately and does not require a large amount of data to be optimized for a given form factor, and the network model can be considered as a standalone module which is highly transferable independently from the microphone array. We name the ML module in our solution as GSENet, short for Guided Speech Enhancement Network. We demonstrate its effectiveness on real world data collected on multi-microphone devices in terms of the suppression of noise and interfering speech.

An authentic face restoration system is becoming increasingly demanding in many computer vision applications, e.g., image enhancement, video communication, and taking portrait. Most of the advanced face restoration models can recover high-quality faces from low-quality ones but usually fail to faithfully generate realistic and high-frequency details that are favored by users. To achieve authentic restoration, we propose IDM, an Iteratively learned face restoration system based on denoising Diffusion Models (DDMs). We define the criterion of an authentic face restoration system, and argue that denoising diffusion models are naturally endowed with this property from two aspects: intrinsic iterative refinement and extrinsic iterative enhancement. Intrinsic learning can preserve the content well and gradually refine the high-quality details, while extrinsic enhancement helps clean the data and improve the restoration task one step further. We demonstrate superior performance on blind face restoration tasks. Beyond restoration, we find the authentically cleaned data by the proposed restoration system is also helpful to image generation tasks in terms of training stabilization and sample quality. Without modifying the models, we achieve better quality than state-of-the-art on FFHQ and ImageNet generation using either GANs or diffusion models.

Despite the proliferation of generative models, achieving fast sampling during inference without compromising sample diversity and quality remains challenging. Existing models such as Denoising Diffusion Probabilistic Models (DDPM) deliver high-quality, diverse samples but are slowed by an inherently high number of iterative steps. The Denoising Diffusion Generative Adversarial Networks (DDGAN) attempted to circumvent this limitation by integrating a GAN model for larger jumps in the diffusion process. However, DDGAN encountered scalability limitations when applied to large datasets. To address these limitations, we introduce a novel approach that tackles the problem by matching implicit and explicit factors. More specifically, our approach involves utilizing an implicit model to match the marginal distributions of noisy data and the explicit conditional distribution of the forward diffusion. This combination allows us to effectively match the joint denoising distributions. Unlike DDPM but similar to DDGAN, we do not enforce a parametric distribution for the reverse step, enabling us to take large steps during inference. Similar to the DDPM but unlike DDGAN, we take advantage of the exact form of the diffusion process. We demonstrate that our proposed method obtains comparable generative performance to diffusion-based models and vastly superior results to models with a small number of sampling steps.

3D object detection has recently become popular due to many applications in robotics, augmented reality, autonomy, and image retrieval. We introduce the Objectron dataset to advance the state of the art in 3D object detection and foster new research and applications, such as 3D object tracking, view synthesis, and improved 3D shape representation. The dataset contains object-centric short videos with pose annotations for nine categories and includes 4 million annotated images in 14, 819 annotated videos. We also propose a new evaluation metric, 3D Intersection over Union, for 3D object detection. We demonstrate the usefulness of our dataset in 3D object detection and novel view synthesis tasks by providing baseline models trained on this dataset. Our dataset and evaluation source code are available online at Github.com/google-research-datasets/Objectron.

We present an algorithm for finding temporally consistent occlusion boundaries in videos to support segmentation of dynamic scenes. We learn occlusion boundaries in a pair wise Markov random field (MRF) framework. We first estimate the probability of an spatio-temporal edge being an occlusion boundary by using appearance, flow, and geometric features. Next, we enforce occlusion boundary continuity in a MRF model by learning pair wise occlusion probabilities using a random forest. Then, we temporally smooth boundaries to remove temporal inconsistencies in occlusion boundary estimation. Our proposed framework provides an efficient approach for finding temporally consistent occlusion boundaries in video by utilizing causality, redundancy in videos, and semantic layout of the scene. We have developed a dataset with fully annotated ground-truth occlusion boundaries of over 30 videos (~5000 frames). This dataset is used to evaluate temporal occlusion boundaries and provides a much needed baseline for future studies. We perform experiments to demonstrate the role of scene layout, and temporal information for occlusion reasoning in dynamic scenes.

We present a novel algorithm for estimating the broad 3D geometric structure of outdoor video scenes. Leveraging spatio-temporal video segmentation, we decompose a dynamic scene captured by a video into geometric classes, based on predictions made by region-classifiers that are trained on appearance and motion features. By examining the homogeneity of the prediction, we combine predictions across multiple segmentation hierarchy levels alleviating the need to determine the granularity a priori. We built a novel, extensive dataset on geometric context of video to evaluate our method, consisting of over 100 ground-truth annotated outdoor videos with over 20,000 frames. To further scale beyond this dataset, we propose a semi-supervised learning framework to expand the pool of labeled data with high confidence predictions obtained from unlabeled data. Our system produces an accurate prediction of geometric context of video achieving 96% accuracy across main geometric classes.

We present a novel data-driven technique for radiometric self-calibration of video from an unknown camera. Our approach self-calibrates radiometric variations in video, and is applied as a post-process; there is no need to access the camera, and in particular it is applicable to internet videos. This technique builds on empirical evidence that in video the camera response function (CRF) should be regarded time variant, as it changes with scene content and exposure, instead of relying on a single camera response function. We show that a time-varying mixture of responses produces better accuracy and consistently reduces the error in mapping intensity to irradiance when compared to a single response model. Furthermore, our mixture model counteracts the effects of possible nonlinear exposure-dependent intensity perturbations and white-balance changes caused by proprietary camera firmware. We further show how radiometrically calibrated video improves the performance of other video analysis algorithms, enabling a video segmentation algorithm to be invariant to exposure and gain variations over the sequence. We validate our data-driven technique on videos from a variety of cameras and demonstrate the generality of our approach by applying it to internet video.

We present a novel algorithm for efficient removal of rolling shutter distortions in uncalibrated streaming videos. Our proposed method is calibration free as it does not need any knowledge of the camera used, nor does it require calibration using specially recorded calibration sequences. Our algorithm can perform rolling shutter removal under varying focal lengths, as in videos from CMOS cameras equipped with an optical zoom. We evaluate our approach across a broad range of cameras and video sequences demonstrating robustness, scaleability, and repeatability. We also conducted a user study, which demonstrates preference for the output of our algorithm over other state-of-the art methods. Our algorithm is computationally efficient, easy to parallelize, and robust to challenging artifacts introduced by various cameras with differing technologies.

We present a novel algorithm for automatically applying constrainable, L1-optimal camera paths to generate stabilized videos by removing undesired motions. Our goal is to compute camera paths that are composed of constant, linear and parabolic segments mimicking the camera motions employed by professional cinematographers. To this end, our algorithm is based on a linear programming framework to minimize the first, second, and third derivatives of the resulting camera path. Our method allows for video stabilization beyond the conventional filtering of camera paths that only suppresses high frequency jitter. We incorporate additional constraints on the path of the camera directly in our algorithm, allowing for stabilized and retargeted videos. Our approach accomplishes this without the need of user interaction or costly 3D reconstruction of the scene, and works as a post-process for videos from any camera or from an online source.

We introduce a new algorithm for video retargeting that uses discontinuous seam-carving in both space and time for resizing videos. Our algorithm relies on a novel appearance-based temporal coherence formulation that allows for frame-by-frame processing and results in temporally discontinuous seams, as opposed to geometrically smooth and continuous seams. This formulation optimizes the difference in appearance of the resultant retargeted frame to the optimal temporally coherent one, and allows for carving around fast moving salient regions. Additionally, we generalize the idea of appearance-based coherence to the spatial domain by introducing piece-wise spatial seams. Our spatial coherence measure minimizes the change in gradients during retargeting, which preserves spatial detail better than minimization of color difference alone. We also show that per-frame saliency (gradient-based or feature-based) does not always produce desirable retargeting results and propose a novel automatically computed measure of spatio-temporal saliency. As needed, a user may also augment the saliency by interactive region-brushing. Our retargeting algorithm processes the video sequentially, making it conducive for streaming applications.

We present an efficient and scalable technique for spatiotemporal segmentation of long video sequences using a hierarchical graph-based algorithm. We begin by over-segmenting a volumetric video graph into space-time regions grouped by appearance. We then construct a “region graph” over the obtained segmentation and iteratively repeat this process over multiple levels to create a tree of spatio-temporal segmentations. This hierarchical approach generates high quality segmentations, which are temporally coherent with stable region boundaries, and allows subsequent applications to choose from varying levels of granularity. We further improve segmentation quality by using dense optical flow to guide temporal connections in the initial graph. We also propose two novel approaches to improve the scalability of our technique: (a) a parallel out-of-core algorithm that can process volumes much larger than an in-core algorithm, and (b) a clip-based processing algorithm that divides the video into overlapping clips in time, and segments them successively while enforcing consistency. We demonstrate hierarchical segmentations on video shots as long as 40 seconds, and even support a streaming mode for arbitrarily long videos, albeit without the ability to process them hierarchically.

We present a novel approach for robust localization of multiple people observed using multiple cameras. We use this location information to generate sports visualizations, which include displaying a virtual offside line in soccer games, and showing players' positions and motion patterns. Our main contribution is the modeling and analysis for the problem of fusing corresponding players' positional information as finding minimum weight K-length cycles in complete K-partite graphs. To this end, we use a dynamic programming based approach that varies over a continuum of being maximally to minimally greedy in terms of the number of paths explored at each iteration. We present an end-to-end sports visualization framework that employs our proposed algorithm-class. We demonstrate the robustness of our framework by testing it on 60,000 frames of soccer footage captured over 5 different illumination conditions, play types, and team attire.

Videos of multi-player team sports provide a challenging domain for dynamic scene analysis. Player actions and interactions are complex as they are driven by many factors, such as the short-term goals of the individual player, the overall team strategy, the rules of the sport, and the current context of the game. We show that constrained multi-agent events can be analyzed and even predicted from video. Such analysis requires estimating the global movements of all players in the scene at any time, and is needed for modeling and predicting how the multi-agent play evolves over time on the field. To this end, we propose a novel approach to detect the locations of where the play evolution will proceed, e.g. where interesting events will occur, by tracking player positions and movements over time. We start by extracting the ground level sparse movement of players in each time-step, and then generate a dense motion field. Using this field we detect locations where the motion converges, implying positions towards which the play is evolving. We evaluate our approach by analyzing videos of a variety of complex soccer plays.

We propose the use of 3D (2D+time) Shape Context to recognize the spatial and temporal details inherent in human actions. We represent an action in a video sequence by a 3D point cloud extracted by sampling 2D silhouettes over time. A non-uniform sampling method is introduced that gives preference to fast moving body parts using a Euclidean 3D Distance Transform. Actions are then classified by matching the extracted point clouds. Our proposed approach is based on a global matching and does not require specific training to learn the model. We test the approach thoroughly on two publicly available datasets and compare to several state-of-the-art methods. The achieved classification accuracy is on par with or superior to the best results reported to date.