CVPR 2023 论文列表

The abundance of instructional videos and their narrations over the Internet offers an exciting avenue for understanding procedural activities. In this work, we propose to learn video representation that encodes both action steps and their temporal ordering, based on a large-scale dataset of web instructional videos and their narrations, without using human annotations. Our method jointly learns a video representation to encode individual step concepts, and a deep probabilistic model to capture both temporal dependencies and immense individual variations in the step ordering. We empirically demonstrate that learning temporal ordering not only enables new capabilities for procedure reasoning, but also reinforces the recognition of individual steps. Our model significantly advances the state-of-the-art results on step classification (+2.8% / +3.3% on COIN / EPIC-Kitchens) and step forecasting (+7.4% on COIN). Moreover, our model attains promising results in zero-shot inference for step classification and forecasting, as well as in predicting diverse and plausible steps for incomplete procedures. Our code is available at https://github.com/facebookresearch/ProcedureVRL.

We propose a method to learn a high-quality implicit 3D head avatar from a monocular RGB video captured in the wild. The learnt avatar is driven by a parametric face model to achieve user-controlled facial expressions and head poses. Our hybrid pipeline combines the geometry prior and dynamic tracking of a 3DMM with a neural radiance field to achieve fine-grained control and photorealism. To reduce over-smoothing and improve out-of-model expressions synthesis, we propose to predict local features anchored on the 3DMM geometry. These learnt features are driven by 3DMM deformation and interpolated in 3D space to yield the volumetric radiance at a designated query point. We further show that using a Convolutional Neural Network in the UV space is critical in incorporating spatial context and producing representative local features. Extensive experiments show that we are able to reconstruct high-quality avatars, with more accurate expression-dependent details, good generalization to out-of-training expressions, and quantitatively superior renderings compared to other state-of-the-art approaches.

Visual representation based on covariance matrix has demonstrates its efficacy for image classification by characterising the pairwise correlation of different channels in convolutional feature maps. However, pairwise correlation will become misleading once there is another channel correlating with both channels of interest, resulting in the "confounding" effect. For this case, "partial correlation" which removes the confounding effect shall be estimated instead. Nevertheless, reliably estimating partial correlation requires to solve a symmetric positive definite matrix optimisation, known as sparse inverse covariance estimation (SICE). How to incorporate this process into CNN remains an open issue. In this work, we formulate SICE as a novel structured layer of CNN. To ensure end-to-end trainability, we develop an iterative method to solve the above matrix optimisation during forward and backward propagation steps. Our work obtains a partial correlation based deep visual representation and mitigates the small sample problem often encountered by covariance matrix estimation in CNN. Computationally, our model can be effectively trained with GPU and works well with a large number of channels of advanced CNNs. Experiments show the efficacy and superior classification performance of our deep visual representation compared to covariance matrix based counterparts.

Generalized few-shot semantic segmentation (GFSS) distinguishes pixels of base and novel classes from the background simultaneously, conditioning on sufficient data of base classes and a few examples from novel class. A typical GFSS approach has two training phases: base class learning and novel class updating. Nevertheless, such a stand-alone updating process often compromises the well-learnt features and results in performance drop on base classes. In this paper, we propose a new idea of leveraging Projection onto Orthogonal Prototypes (POP), which updates features to identify novel classes without compromising base classes. POP builds a set of orthogonal prototypes, each of which represents a semantic class, and makes the prediction for each class separately based on the features projected onto its prototype. Technically, POP first learns prototypes on base data, and then extends the prototype set to novel classes. The orthogonal constraint of POP encourages the orthogonality between the learnt prototypes and thus mitigates the influence on base class features when generalizing to novel prototypes. Moreover, we capitalize on the residual of feature projection as the background representation to dynamically fit semantic shifting (i.e., background no longer includes the pixels of novel classes in updating phase). Extensive experiments on two benchmarks demonstrate that our POP achieves superior performances on novel classes without sacrificing much accuracy on base classes. Notably, POP outperforms the state-of-the-art fine-tuning by 3.93% overall mIoU on PASCAL-5i in 5-shot scenario.

This paper address the problem of optical expansion (OE). OE describes the object scale change between two frames, widely used in monocular 3D vision tasks. Previous methods estimate optical expansion mainly from optical flow results, but this two-stage architecture makes their results limited by the accuracy of optical flow and less robust. To solve these problems, we propose the concept of 3D optical flow by integrating optical expansion into the 2D optical flow, which is implemented by a plug-and-play module, namely TPCV. TPCV implements matching features at the correct location and scale, thus allowing the simultaneous optimization of optical flow and optical expansion tasks. Experimentally, we apply TPCV to the RAFT optical flow baseline. Experimental results show that the baseline optical flow performance is substantially improved. Moreover, we apply the optical flow and optical expansion results to various dynamic 3D vision tasks, including motion-in-depth, time-to-collision, and scene flow, often achieving significant improvement over the prior SOTA. Code will be available at https://github.com/HanLingsgjk/TPCV.

In this paper, we consider the problem of open-vocabulary semantic segmentation (OVS), which aims to segment objects of arbitrary classes instead of pre-defined, closed-set categories. The main contributions are as follows: First, we propose a transformer-based model for OVS, termed as OVSegmentor, which only exploits web-crawled image-text pairs for pre-training without using any mask annotations. OVSegmentor assembles the image pixels into a set of learnable group tokens via a slot-attention based binding module, and aligns the group tokens to the corresponding caption embedding. Second, we propose two proxy tasks for training, namely masked entity completion and cross-image mask consistency. The former aims to infer all masked entities in the caption given the group tokens, that enables the model to learn fine-grained alignment between visual groups and text entities. The latter enforces consistent mask predictions between images that contain shared entities, which encourages the model to learn visual invariance. Third, we construct CC4M dataset for pre-training by filtering CC12M with frequently appeared entities, which significantly improves training efficiency. Fourth, we perform zero-shot transfer on three benchmark datasets, PASCAL VOC 2012, PASCAL Context, and COCO Object. Our model achieves superior segmentation results over the state-of-the-art method by using only 3% data (4M vs 134M) for pre-training. Code and pre-trained models will be released for future research.

This paper addresses the challenge of efficiently reconstructing volumetric videos of dynamic humans from sparse multi-view videos. Some recent works represent a dynamic human as a canonical neural radiance field (NeRF) and a motion field, which are learned from input videos through differentiable rendering. But the per-scene optimization generally requires hours. Other generalizable NeRF models leverage learned prior from datasets to reduce the optimization time by only finetuning on new scenes at the cost of visual fidelity. In this paper, we propose a novel method for learning neural volumetric representations of dynamic humans in minutes with competitive visual quality. Specifically, we define a novel part-based voxelized human representation to better distribute the representational power of the network to different human parts. Furthermore, we propose a novel 2D motion parameterization scheme to increase the convergence rate of deformation field learning. Experiments demonstrate that our model can be learned 100 times faster than previous per-scene optimization methods while being competitive in the rendering quality. Training our model on a 512x512 video with 100 frames typically takes about 5 minutes on a single RTX 3090 GPU. The code is available on our project page: https://zju3dv.github.io/instant_nvr

Generative models show good potential for recovering 3D faces beyond limited shape assumptions. While plausible details and resolutions are achieved, these models easily fail under extreme conditions of pose, shadow or appearance, due to the entangled fitting or lack of multi-view priors. To address this problem, this paper presents a novel Neural Proto-face Field (NPF) for unsupervised robust 3D face modeling. Instead of using constrained images as Neural Radiance Field (NeRF), NPF disentangles the common/specific facial cues, i.e., ID, expression and scene-specific details from in-the-wild photo collections. Specifically, NPF learns a face prototype to aggregate 3D-consistent identity via uncertainty modeling, extracting multi-image priors from a photo collection. NPF then learns to deform the prototype with the appropriate facial expressions, constrained by a loss of expression consistency and personal idiosyncrasies. Finally, NPF is optimized to fit a target image in the collection, recovering specific details of appearance and geometry. In this way, the generative model benefits from multi-image priors and meaningful facial structures. Extensive experiments on benchmarks show that NPF recovers superior or competitive facial shapes and textures, compared to state-of-the-art methods.

We propose a novel 3D morphable model for complete human heads based on hybrid neural fields. At the core of our model lies a neural parametric representation that disentangles identity and expressions in disjoint latent spaces. To this end, we capture a person's identity in a canonical space as a signed distance field (SDF), and model facial expressions with a neural deformation field. In addition, our representation achieves high-fidelity local detail by introducing an ensemble of local fields centered around facial anchor points. To facilitate generalization, we train our model on a newly-captured dataset of over 3700 head scans from 203 different identities using a custom high-end 3D scanning setup. Our dataset significantly exceeds comparable existing datasets, both with respect to quality and completeness of geometry, averaging around 3.5M mesh faces per scan. Finally, we demonstrate that our approach outperforms state-of-the-art methods in terms of fitting error and reconstruction quality.

Neural radiance fields (NeRFs) enable novel view synthesis with unprecedented visual quality. However, to render photorealistic images, NeRFs require hundreds of deep multilayer perceptron (MLP) evaluations -- for each pixel. This is prohibitively expensive and makes real-time rendering infeasible, even on powerful modern GPUs. In this paper, we propose a novel approach to distill and bake NeRFs into highly efficient mesh-based neural representations that are fully compatible with the massively parallel graphics rendering pipeline. We represent scenes as neural radiance features encoded on a two-layer duplex mesh, which effectively overcomes the inherent inaccuracies in 3D surface reconstruction by learning the aggregated radiance information from a reliable interval of ray-surface intersections. To exploit local geometric relationships of nearby pixels, we leverage screen-space convolutions instead of the MLPs used in NeRFs to achieve high-quality appearance. Finally, the performance of the whole framework is further boosted by a novel multi-view distillation optimization strategy. We demonstrate the effectiveness and superiority of our approach via extensive experiments on a range of standard datasets.

Weakly supervised dense object localization (WSDOL) relies generally on Class Activation Mapping (CAM), which exploits the correlation between the class weights of the image classifier and the pixel-level features. Due to the limited ability to address intra-class variations, the image classifier cannot properly associate the pixel features, leading to inaccurate dense localization maps. In this paper, we propose to explicitly construct multi-modal class representations by leveraging the Contrastive Language-Image Pre-training (CLIP), to guide dense localization. More specifically, we propose a unified transformer framework to learn two-modalities of class-specific tokens, i.e., class-specific visual and textual tokens. The former captures semantics from the target visual data while the latter exploits the class-related language priors from CLIP, providing complementary information to better perceive the intra-class diversities. In addition, we propose to enrich the multi-modal class-specific tokens with sample-specific contexts comprising visual context and image-language context. This enables more adaptive class representation learning, which further facilitates dense localization. Extensive experiments show the superiority of the proposed method for WSDOL on two multi-label datasets, i.e., PASCAL VOC and MS COCO, and one single-label dataset, i.e., OpenImages. Our dense localization maps also lead to the state-of-the-art weakly supervised semantic segmentation (WSSS) results on PASCAL VOC and MS COCO.

In this paper, we propose a novel hybrid representation and end-to-end trainable network architecture to model fully editable and customizable neural avatars. At the core of our work lies a representation that combines the modeling power of neural fields with the ease of use and inherent 3D consistency of skinned meshes. To this end, we construct a trainable feature codebook to store local geometry and texture features on the vertices of a deformable body model, thus exploiting its consistent topology under articulation. This representation is then employed in a generative auto-decoder architecture that admits fitting to unseen scans and sampling of realistic avatars with varied appearances and geometries. Furthermore, our representation allows local editing by swapping local features between 3D assets. To verify our method for avatar creation and editing, we contribute a new high-quality dataset, dubbed CustomHumans, for training and evaluation. Our experiments quantitatively and qualitatively show that our method generates diverse detailed avatars and achieves better model fitting performance compared to state-of-the-art methods. Our code and dataset are available at https://ait.ethz.ch/custom-humans.

This paper studies the fundamental problem of learning multi-layer generator models. The multi-layer generator model builds multiple layers of latent variables as a prior model on top of the generator, which benefits learning complex data distribution and hierarchical representations. However, such a prior model usually focuses on modeling inter-layer relations between latent variables by assuming non-informative (conditional) Gaussian distributions, which can be limited in model expressivity. To tackle this issue and learn more expressive prior models, we propose an energy-based model (EBM) on the joint latent space over all layers of latent variables with the multi-layer generator as its backbone. Such joint latent space EBM prior model captures the intra-layer contextual relations at each layer through layer-wise energy terms, and latent variables across different layers are jointly corrected. We develop a joint training scheme via maximum likelihood estimation (MLE), which involves Markov Chain Monte Carlo (MCMC) sampling for both prior and posterior distributions of the latent variables from different layers. To ensure efficient inference and learning, we further propose a variational training scheme where an inference model is used to amortize the costly posterior MCMC sampling. Our experiments demonstrate that the learned model can be expressive in generating high-quality images and capturing hierarchical features for better outlier detection.

This paper aims to establish a generic multi-modal foundation model that has the scalable capability to massive downstream applications in E-commerce. Recently, large-scale vision-language pretraining approaches have achieved remarkable advances in the general domain. However, due to the significant differences between natural and product images, directly applying these frameworks for modeling image-level representations to E-commerce will be inevitably sub-optimal. To this end, we propose an instance-centric multi-modal pretraining paradigm called ECLIP in this work. In detail, we craft a decoder architecture that introduces a set of learnable instance queries to explicitly aggregate instance-level semantics. Moreover, to enable the model to focus on the desired product instance without reliance on expensive manual annotations, two specially configured pretext tasks are further proposed. Pretrained on the 100 million E-commerce-related data, ECLIP successfully extracts more generic, semantic-rich, and robust representations. Extensive experimental results show that, without further fine-tuning, ECLIP surpasses existing methods by a large margin on a broad range of downstream tasks, demonstrating the strong transferability to real-world E-commerce applications.

The real-world data tends to be heavily imbalanced and severely skew the data-driven deep neural networks, which makes Long-Tailed Recognition (LTR) a massive challenging task. Existing LTR methods seldom train Vision Transformers (ViTs) with Long-Tailed (LT) data, while the off-the-shelf pretrain weight of ViTs always leads to unfair comparisons. In this paper, we systematically investigate the ViTs' performance in LTR and propose LiVT to train ViTs from scratch only with LT data. With the observation that ViTs suffer more severe LTR problems, we conduct Masked Generative Pretraining (MGP) to learn generalized features. With ample and solid evidence, we show that MGP is more robust than supervised manners. Although Binary Cross Entropy (BCE) loss performs well with ViTs, it struggles on the LTR tasks. We further propose the balanced BCE to ameliorate it with strong theoretical groundings. Specially, we derive the unbiased extension of Sigmoid and compensate extra logit margins for deploying it. Our Bal-BCE contributes to the quick convergence of ViTs in just a few epochs. Extensive experiments demonstrate that with MGP and Bal-BCE, LiVT successfully trains ViTs well without any additional data and outperforms comparable state-of-the-art methods significantly, e.g., our ViT-B achieves 81.0% Top-1 accuracy in iNaturalist 2018 without bells and whistles. Code is available at https://github.com/XuZhengzhuo/LiVT.

We propose the first framework to learn control policies for vision-based human-to-robot handovers, a critical task for human-robot interaction. While research in Embodied AI has made significant progress in training robot agents in simulated environments, interacting with humans remains challenging due to the difficulties of simulating humans. Fortunately, recent research has developed realistic simulated environments for human-to-robot handovers. Leveraging this result, we introduce a method that is trained with a human-in-the-loop via a two-stage teacher-student framework that uses motion and grasp planning, reinforcement learning, and self-supervision. We show significant performance gains over baselines on a simulation benchmark, sim-to-sim transfer and sim-to-real transfer.

We present a novel method for recovering the absolute pose and shape of a human in a pre-scanned scene given a single image. Unlike previous methods that perform sceneaware mesh optimization, we propose to first estimate absolute position and dense scene contacts with a sparse 3D CNN, and later enhance a pretrained human mesh recovery network by cross-attention with the derived 3D scene cues. Joint learning on images and scene geometry enables our method to reduce the ambiguity caused by depth and occlusion, resulting in more reasonable global postures and contacts. Encoding scene-aware cues in the network also allows the proposed method to be optimization-free, and opens up the opportunity for real-time applications. The experiments show that the proposed network is capable of recovering accurate and physically-plausible meshes by a single forward pass and outperforms state-of-the-art methods in terms of both accuracy and speed. Code is available on our project page: https://zju3dv.github.io/sahmr/.

Understanding geometric concepts, such as distance and shape, is essential for understanding the real world and also for many vision tasks. To incorporate such information into a visual representation of a scene, we propose learning to represent the scene by sketching, inspired by human behavior. Our method, coined Learning by Sketching (LBS), learns to convert an image into a set of colored strokes that explicitly incorporate the geometric information of the scene in a single inference step without requiring a sketch dataset. A sketch is then generated from the strokes where CLIP-based perceptual loss maintains a semantic similarity between the sketch and the image. We show theoretically that sketching is equivariant with respect to arbitrary affine transformations and thus provably preserves geometric information. Experimental results show that LBS substantially improves the performance of object attribute classification on the unlabeled CLEVR dataset, domain transfer between CLEVR and STL-10 datasets, and for diverse downstream tasks, confirming that LBS provides rich geometric information.

Cross-modal training using 2D-3D paired datasets, such as those containing multi-view images and 3D scene scans, presents an effective way to enhance 2D scene understanding by introducing geometric and view-invariance priors into 2D features. However, the need for large-scale scene datasets can impede scalability and further improvements. This paper explores an alternative learning method by leveraging a lightweight and publicly available type of 3D data in the form of CAD models. We construct a 3D space with geometric-aware alignment where the similarity in this space reflects the geometric similarity of CAD models based on the Chamfer distance. The acquired geometric-aware properties are then induced into 2D features, which boost performance on downstream tasks more effectively than existing RGB-CAD approaches. Our technique is not limited to paired RGB-CAD datasets. By training exclusively on pseudo pairs generated from CAD-based reconstruction methods, we enhance the performance of SOTA 2D pre-trained models that use ResNet-50 or ViT-B backbones on various 2D understanding tasks. We also achieve comparable results to SOTA methods trained on scene scans on four tasks in NYUv2, SUNRGB-D, indoor ADE20k, and indoor/outdoor COCO, despite using lightweight CAD models or pseudo data.