Recent advances in foundation models have enabled conversational agents that aim for sustained companionship rather than mere task completion. Yet most still remain unable to support natural, long-term companion-like interactions, resulting in experiences that feel episodic and inauthentic. We argue that current agents overlooked cross-temporal modeling of agents’ social behaviors and internal emotions: generated behaviors rarely influence an agent’s emotional state, and emotional states seldom shape subsequent behaviors. We present Cross-Temporal Emotion Modeling (CTEM), a framework that links long-term behavioral history to moment-to-moment emotional expression. CTEM establishes a closed loop where past experiences update an evolving emotional state; this state conditions immediate interactions; and user feedback continually revises both memory and emotional state, enabling reflection and anticipation. We instantiate CTEM as Auri, a companion agent on an instant-messaging platform, and report a 21-day in-the-wild study showing that CTEM shows improvements in perceived naturalness, coherence, and emotional harmony.

Video object removal aims at erasing a target object in the entire video and filling holes with plausible contents, given an object mask in the first frame as input. Existing solutions mostly break down the task into (supervised) mask tracking and (self-supervised) video completion, and then separately tackle them with tailored designs. In this paper, we introduce a new setup, coined as unified video object removal, where mask tracking and completion are addressed within a unified framework. Despite introducing more challenges, the setup is promising for future practical usage. We embrace the observation that these two sub-tasks have strong inherent connections in terms of pixel-level temporal correspondence. Making full use of the connections could be beneficial considering the complexity of both algorithm and deployment. We propose a single network linking the two sub-tasks by inferring temporal correspondences across multiple frames, i.e., correspondences between valid-valid (V-V) pixel pairs for mask tracking and correspondences between valid-hole (V-H) pixel pairs for video completion. Thanks to the unified setup, the network can be learned end-to-end in a totally unsupervised fashion without any annotations. We demonstrate that our method can generate visually pleasing results and perform favorably against existing separate solutions in realistic test cases.

We propose a novel and general framework to disentangle video data into its dynamic motion and static content components. Our proposed method is a self-supervised pipeline with less assumptions and inductive biases than previous works: it utilizes a transformer-based architecture to jointly generate flexible implicit features for frame-wise motion and clip-wise content, and incorporates a low-bitrate vector quantization as an information bottleneck to promote disentanglement and form a meaningful discrete motion space. The bitrate-controlled latent motion and content are used as conditional inputs to a denoising diffusion model to facilitate self-supervised representation learning. We validate our disentangled representation learning framework on real-world talking head videos with motion transfer and auto-regressive motion generation tasks. Furthermore, we also show that our method can generalize to other types of video data, such as pixel sprites of 2D cartoon characters. Our work presents a new perspective on self-supervised learning of disentangled video representations, contributing to the broader field of video analysis and generation.

The Multiplane Image (MPI), containing a set of fronto-parallel$RGB_{\alpha}$layers, is an effective and efficient representation for view synthesis from sparse inputs. Yet, its fixed structure limits the performance, especially for surfaces imaged at oblique angles. We introduce the Structural MPI (S-MPI), where the plane structure approximates 3D scenes concisely. Conveying$RGB_{\alpha}$contexts with geometrically-faithful structures, the S-MPI directly bridges view synthe-sis and 3D reconstruction. It can not only overcome the critical limitations of MPI, i.e., discretization artifacts from sloped surfaces and abuse of redundant layers, and can also acquire planar 3D reconstruction. Despite the intu-ition and demand of applying S-MPI, great challenges are introduced,$e.g$., high-fidelity approximation for both$RGB_{\alpha}$layers and plane poses, multi-view consistency, non-planar regions modeling, and efficient rendering with intersected planes. Accordingly, we propose a transformer-based network based on a segmentation model [4]. It predicts compact and expressive S-MPI layers with their corresponding masks, poses, and$RGB_{\alpha}$contexts. Non-planar regions are inclusively handled as a special case in our unified frame-work. Multi-view consistency is ensured by sharing global proxy embeddings, which encode plane-level features cov-ering the complete 3D scenes with aligned coordinates. In-tensive experiments show that our method outperforms both previous state-of-the-art MPI-based view synthesis methods and planar reconstruction methods.

Talking head generation is to generate video based on a given source identity and target motion. However, current methods face several challenges that limit the quality and controllability of the generated videos. First, the generated face often has unexpected deformation and severe distortions. Second, the driving image does not explicitly disentangle movement-relevant information, such as poses and expressions, which restricts the manipulation of different attributes during generation. Third, the generated videos tend to have flickering artifacts due to the inconsistency of the extracted landmarks between adjacent frames. In this paper, we propose a novel model that produces high-fidelity talking head videos with free control over head pose and expression. Our method leverages both self-supervised learned landmarks and 3D face model-based landmarks to model the motion. We also introduce a novel motion-aware multi-scale feature alignment module to effectively transfer the motion without face distortion. Furthermore, we enhance the smoothness of the synthesized talking head videos with a feature context adaptation and propagation module. We evaluate our model on challenging datasets and demonstrate its state-of-the-art performance. More information is available at https://yuegao.me/PECHead.

3D shape completion from point clouds is a challenging task, especially from scans of real-world objects. Considering the paucity of 3D shape ground truths for real scans, existing works mainly focus on benchmarking this task on synthetic data, e.g. 3D computer-aided design models. However, the domain gap between synthetic and real data limits the generalizability of these methods. Thus, we propose a new task, SCoDA, for the domain adaptation of real scan shape completion from synthetic data. A new dataset, ScanSalon, is contributed with a bunch of elaborate 3D models created by skillful artists according to scans. To address this new task, we propose a novel cross-domain feature fusion method for knowledge transfer and a novel volume-consistent self-training framework for robust learning from real data. Extensive experiments prove our method is effective to bring an improvement of 6%~7% mIoU.

Previous works on video object segmentation (VOS) are trained on densely annotated videos. Nevertheless, acquiring annotations in pixel level is expensive and time-consuming. In this work, we demonstrate the feasibility of training a satisfactory VOS model on sparsely annotated videos—we merely require two labeled frames per training video while the performance is sustained. We term this novel training paradigm as two-shot video object segmentation, or two-shot VOS for short. The underlying idea is to generate pseudo labels for unlabeled frames during training and to optimize the model on the combination of labeled and pseudo-labeled data. Our approach is extremely simple and can be applied to a majority of existing frameworks. We first pre-train a VOS model on sparsely annotated videos in a semi-supervised manner, with the first frame always being a labeled one. Then, we adopt the pretrained VOS model to generate pseudo labels for all unlabeled frames, which are subsequently stored in a pseudo-label bank. Finally, we retrain a VOS model on both labeled and pseudo-labeled data without any restrictions on the first frame. For the first time, we present a general way to train VOS models on two-shot VOS datasets. By using 7.3% and 2.9% labeled data of YouTube-VOS and DAVIS benchmarks, our approach achieves comparable results in contrast to the counterparts trained on fully labeled set. Code and models are available at https://github.com/yk-pku/Two-shot-Video-Object-Segmentation.

Referring Video Object Segmentation (R-VOS) is a challenging task that aims to segment an object in a video based on a linguistic expression. Most existing R-VOS methods have a critical assumption: the object referred to must appear in the video. This assumption, which we refer to as "semantic consensus", is often violated in real-world scenarios, where the expression may be queried against false videos. In this work, we highlight the need for a robust R-VOS model that can handle semantic mismatches. Accordingly, we propose an extended task called Robust R-VOS (R2-VOS), which accepts unpaired video-text inputs. We tackle this problem by jointly modeling the primary R-VOS problem and its dual (text reconstruction). A structural text-to-text cycle constraint is introduced to discriminate semantic consensus between video-text pairs and impose it in positive pairs, thereby achieving multi-modal alignment from both positive and negative pairs. Our structural constraint effectively addresses the challenge posed by linguistic diversity, overcoming the limitations of previous methods that relied on the point-wise constraint. A new evaluation dataset, R2-Youtube-VOS is constructed to measure the model robustness. Our model achieves state-of-the-art performance on R-VOS benchmarks, Ref-DAVIS17 and Ref-Youtube-VOS, and also our R2-Youtube-VOS dataset.

Recent work on Neural Radiance Fields (NeRF) has demonstrated significant advances in high-quality view synthesis. A major limitation of NeRF is its low rendering efficiency due to the need for multiple network forwardings to render a single pixel. Existing methods to improve NeRF either reduce the number of required samples or optimize the implementation to accelerate the network forwarding. Despite these efforts, the problem of multiple sampling persists due to the intrinsic representation of radiance fields. In contrast, Neural Light Fields (NeLF) reduce the computation cost of NeRF by querying only one single network forwarding per pixel. To achieve a close visual quality to NeRF, existing NeLF methods require significantly larger network capacities which limits their rendering efficiency in practice. In this work, we propose a new representation called Neural Radiance Distribution Field (NeRDF) that targets efficient view synthesis in real-time. Specifically, we use a small network similar to NeRF while preserving the rendering speed with a single network forwarding per pixel as in NeLF. The key is to model the radiance distribution along each ray with frequency basis and predict frequency weights using the network. Pixel values are then computed via volume rendering on radiance distributions. Experiments show that our proposed method offers a better trade-off among speed, quality, and network size than existing methods: we achieve a ∼254× speed-up over NeRF with similar network size, with only a marginal performance decline. Our project page is at yushuang-wu.github.io/NeRDF.

Instance segmentation is a challenging task aiming at classifying and segmenting all object instances of specific classes. While two-stage box-based methods achieve top performances in the image domain, they cannot easily extend their superiority into the video domain. This is because they usually deal with features or images cropped from the detected bounding boxes without alignment, failing to capture pixel-level temporal consistency. We embrace the observation that bottom-up methods dealing with box-free features could offer accurate spacial correlations across frames, which can be fully utilized for object and pixel level tracking. We first propose our bottom-up framework equipped with a temporal context fusion module to better encode inter-frame correlations. Intra-frame cues for semantic segmentation and object localization are simultaneously extracted and reconstructed by corresponding decoders after a shared backbone. For efficient and robust tracking among instances, we introduce an instance-level correspondence across adjacent frames, which is represented by a center-to-center flow, termed as instance flow, to assemble messy dense temporal correspondences. Experiments demonstrate that the proposed method outperforms the state-of-the-art online methods (taking image-level input) on the challenging Youtube-VIS dataset (Yang et al., 2019).

Recently, transformer-based image segmentation methods have achieved notable success against previous solutions. While for video domains, how to effectively model temporal context with the attention of object instances across frames remains an open problem. In this paper, we propose an online video instance segmentation framework with a novel instance-aware temporal fusion method. We first leverage the representation, \ie, a latent code in the global context (instance code) and CNN feature maps to represent instance- and pixel-level features. Based on this representation, we introduce a cropping-free temporal fusion approach to model the temporal consistency between video frames. Specifically, we encode global instance-specific information in the instance code and build up inter-frame contextual fusion with hybrid attentions between the instance codes and CNN feature maps. Inter-frame consistency between the instance codes is further enforced with order constraints. By leveraging the learned hybrid temporal consistency, we are able to directly retrieve and maintain instance identities across frames, eliminating the complicated frame-wise instance matching in prior methods. Extensive experiments have been conducted on popular VIS datasets, i.e. Youtube-VIS-19/21. Our model achieves the best performance among all online VIS methods. Notably, our model also eclipses all offline methods when using the ResNet-50 backbone.

Error propagation is a general but crucial problem in online semi-supervised video object segmentation. We aim to suppress error propagation through a correction mechanism with high reliability. The key insight is to disentangle the correction from the conventional mask propagation process with reliable cues. We introduce two modulators, propagation and correction modulators, to separately perform channel-wise recalibration on the target frame embeddings according to local temporal correlations and reliable references respectively. Specifically, we assemble the modulators with a cascaded propagation-correction scheme. This avoids overriding the effects of the reliable correction modulator by the propagation modulator. Although the reference frame with the ground truth label provides reliable cues, it could be very different from the target frame and introduce uncertain or incomplete correlations. We augment the reference cues by supplementing reliable feature patches to a maintained pool, thus offering more comprehensive and expressive object representations to the modulators. In addition, a reliability filter is designed to retrieve reliable patches and pass them in subsequent frames. Our model achieves state-of-the-art performance on YouTube-VOS18, YouTube-VOS19 and DAVIS17-Val/Test benchmarks. Extensive experiments demonstrate that the correction mechanism provides considerable performance gain by fully utilizing reliable guidance.

Multimodal and multi-domain stylization are two important problems in the field of image style transfer. Currently, there are few methods that can perform multimodal and multi-domain stylization simultaneously. In this study, we propose a unified framework for multimodal and multi-domain style transfer with the support of both exemplar-based reference and randomly sampled guidance. The key component of our method is a novel style distribution alignment module that eliminates the explicit distribution gaps between various style domains and reduces the risk of mode collapse. The multimodal diversity is ensured by either guidance from multiple images or random style codes, while the multi-domain controllability is directly achieved by using a domain label. We validate our proposed framework on painting style transfer with various artistic styles and genres. Qualitative and quantitative comparisons with state-of-the-art methods demonstrate that our method can generate high-quality results of multi-domain styles and multimodal instances from reference style guidance or a random sampled style.

We present a new weakly supervised learning-based method for generating novel category-specific 3D shapes from unoccluded image collections. Our method is weakly supervised and only requires silhouette annotations from unoccluded, category-specific objects. Our method does not require access to the object’s 3D shape, multiple observations per object from different views, intra-image pixel correspondences, or any view annotations. Key to our method is a novel multi-projection generative adversarial network (MP-GAN) that trains a 3D shape generator to be consistent with multiple 2D projections of the 3D shapes, and without direct access to these 3D shapes. This is achieved through multiple discriminators that encode the distribution of 2D projections of the 3D shapes seen from a different views. Additionally, to determine the view information for each silhouette image, we also train a view prediction network on visualizations of 3D shapes synthesized by the generator. We iteratively alternate between training the generator and training the view prediction network. We validate our multi-projection GAN on both synthetic and real image datasets. Furthermore, we also show that multi-projection GANs can aid in learning other high-dimensional distributions from lower dimensional training datasets, such as material-class specific spatially varying reflectance properties from images.

In this paper we present a unified deep inverse rendering framework for estimating the spatially-varying appearance properties of a planar exemplar from an arbitrary number of input photographs, ranging from just a single photograph to many photographs. The precision of the estimated appearance scales from plausible when the input photographs fails to capture all the reflectance information, to accurate for large input sets. A key distinguishing feature of our framework is that it directly optimizes for the appearance parameters in a latent embedded space of spatially-varying appearance, such that no handcrafted heuristics are needed to regularize the optimization. This latent embedding is learned through a fully convolutional auto-encoder that has been designed to regularize the optimization. Our framework not only supports an arbitrary number of input photographs, but also at high resolution. We demonstrate and evaluate our deep inverse rendering solution on a wide variety of publicly available datasets.

Abstract This paper presents a deep learning based method for estimating the spatially varying surface reflectance properties from a single image of a planar surface under unknown natural lighting trained using only photographs of exemplar materials without referencing any artist generated or densely measured spatially varying surface reflectance training data. Our method is based on an empirical study of Li et al.'s [ LDPT17 ] self‐augmentation training strategy that shows that the main role of the initial approximative network is to provide guidance on the inherent ambiguities in single image appearance estimation. Furthermore, our study indicates that this initial network can be inexact (i.e., trained from other data sources) as long as it resolves the inherent ambiguities. We show that the single image estimation network trained without manually labeled data outperforms prior work in terms of accuracy as well as generality.

We present a convolutional neural network (CNN) based solution for modeling physically plausible spatially varying surface reflectance functions (SVBRDF) from a single photograph of a planar material sample under unknown natural illumination. Gathering a sufficiently large set of labeled training pairs consisting of photographs of SVBRDF samples and corresponding reflectance parameters, is a difficult and arduous process. To reduce the amount of required labeled training data, we propose to leverage the appearance information embedded in unlabeled images of spatially varying materials to self-augment the training process. Starting from an initial approximative network obtained from a small set of labeled training pairs, we estimate provisional model parameters for each unlabeled training exemplar. Given this provisional reflectance estimate, we then synthesize a novel temporarylabeledtraining pair by rendering the exact corresponding image under a new lighting condition. After refining the network using these additional training samples, we re-estimate the provisional model parameters for the unlabeled data and repeat the self-augmentation process until convergence. We demonstrate the efficacy of the proposed network structure on spatially varying wood, metals, and plastics, as well as thoroughly validate the effectiveness of the self-augmentation training process.