CVPR 2023 Papers

In this work, we tackle the challenging problem of learning-based single-view 3D hair modeling. Due to the great difficulty of collecting paired real image and 3D hair data, using synthetic data to provide prior knowledge for real domain becomes a leading solution. This unfortunately introduces the challenge of domain gap. Due to the inherent difficulty of realistic hair rendering, existing methods typically use orientation maps instead of hair images as input to bridge the gap. We firmly think an intermediate representation is essential, but we argue that orientation map using the dominant filtering-based methods is sensitive to uncertain noise and far from a competent representation. Thus, we first raise this issue up and propose a novel intermediate representation, termed as HairStep, which consists of a strand map and a depth map. It is found that HairStep not only provides sufficient information for accurate 3D hair modeling, but also is feasible to be inferred from real images. Specifically, we collect a dataset of 1,250 portrait images with two types of annotations. A learning framework is further designed to transfer real images to the strand map and depth map. It is noted that, an extra bonus of our new dataset is the first quantitative metric for 3D hair modeling. Our experiments show that HairStep narrows the domain gap between synthetic and real and achieves state-of-the-art performance on single-view 3D hair reconstruction.

We present the Habitat-Matterport 3D Semantics (HM3DSEM) dataset. HM3DSEM is the largest dataset of 3D real-world spaces with densely annotated semantics that is currently available to the academic community. It consists of 142,646 object instance annotations across 216 3D spaces and 3,100 rooms within those spaces. The scale, quality, and diversity of object annotations far exceed those of prior datasets. A key difference setting apart HM3DSEM from other datasets is the use of texture information to annotate pixel-accurate object boundaries. We demonstrate the effectiveness of HM3DSEM dataset for the Object Goal Navigation task using different methods. Policies trained using HM3DSEM perform outperform those trained on prior datasets. Introduction of HM3DSEM in the Habitat ObjectNav Challenge lead to an increase in participation from 400 submissions in 2021 to 1022 submissions in 2022. Project page: https://aihabitat.org/datasets/hm3d-semantics/

Supervised learning of skeleton sequence encoders for action recognition has received significant attention in recent times. However, learning such encoders without labels continues to be a challenging problem. While prior works have shown promising results by applying contrastive learning to pose sequences, the quality of the learned representations is often observed to be closely tied to data augmentations that are used to craft the positives. However, augmenting pose sequences is a difficult task as the geometric constraints among the skeleton joints need to be enforced to make the augmentations realistic for that action. In this work, we propose a new contrastive learning approach to train models for skeleton-based action recognition without labels. Our key contribution is a simple module, HaLP - to Hallucinate Latent Positives for contrastive learning. Specifically, HaLP explores the latent space of poses in suitable directions to generate new positives. To this end, we present a novel optimization formulation to solve for the synthetic positives with an explicit control on their hardness. We propose approximations to the objective, making them solvable in closed form with minimal overhead. We show via experiments that using these generated positives within a standard contrastive learning framework leads to consistent improvements across benchmarks such as NTU-60, NTU-120, and PKU-II on tasks like linear evaluation, transfer learning, and kNN evaluation. Our code can be found at https://github.com/anshulbshah/HaLP.

In this paper, we focus on the problem of category-level object pose estimation, which is challenging due to the large intra-category shape variation. 3D graph convolution (3D-GC) based methods have been widely used to extract local geometric features, but they have limitations for complex shaped objects and are sensitive to noise. Moreover, the scale and translation invariant properties of 3D-GC restrict the perception of an object's size and translation information. In this paper, we propose a simple network structure, the HS-layer, which extends 3D-GC to extract hybrid scope latent features from point cloud data for category-level object pose estimation tasks. The proposed HS-layer: 1) is able to perceive local-global geometric structure and global information, 2) is robust to noise, and 3) can encode size and translation information. Our experiments show that the simple replacement of the 3D-GC layer with the proposed HS-layer on the baseline method (GPV-Pose) achieves a significant improvement, with the performance increased by 14.5% on 5d2cm metric and 10.3% on IoU75. Our method outperforms the state-of-the-art methods by a large margin (8.3% on 5d2cm, 6.9% on IoU75) on REAL275 dataset and runs in real-time (50 FPS).

Depth estimation from a monocular 360 image is a burgeoning problem owing to its holistic sensing of a scene. Recently, some methods, e.g., OmniFusion, have applied the tangent projection (TP) to represent a 360 image and predicted depth values via patch-wise regressions, which are merged to get a depth map with equirectangular projection (ERP) format. However, these methods suffer from 1) non-trivial process of merging a large number of patches; 2) capturing less holistic-with-regional contextual information by directly regressing the depth value of each pixel. In this paper, we propose a novel framework, HRDFuse, that subtly combines the potential of convolutional neural networks (CNNs) and transformers by collaboratively learning the holistic contextual information from the ERP and the regional structural information from the TP. Firstly, we propose a spatial feature alignment (SFA) module that learns feature similarities between the TP and ERP to aggregate the TP features into a complete ERP feature map in a pixel-wise manner. Secondly, we propose a collaborative depth distribution classification (CDDC) module that learns the holistic-with-regional histograms capturing the ERP and TP depth distributions. As such, the final depth values can be predicted as a linear combination of histogram bin centers. Lastly, we adaptively combine the depth predictions from ERP and TP to obtain the final depth map. Extensive experiments show that our method predicts more smooth and accurate depth results while achieving favorably better results than the SOTA methods.

Instead of searching the entire network directly, current NAS approaches increasingly search for multiple relatively small cells to reduce search costs. A major challenge is to jointly measure the similarity of cell micro-architectures and the difference in macro-architectures between different cell-based networks. Recently, optimal transport (OT) has been successfully applied to NAS as it can capture the operational and structural similarity across various networks. However, existing OT-based NAS methods either ignore the cell similarity or focus solely on searching for a single cell architecture. To address these issues, we propose a hierarchical optimal transport metric called HOTNN for measuring the similarity of different networks. In HOTNN, the cell-level similarity computes the OT distance between cells in various networks by considering the similarity of each node and the differences in the information flow costs between node pairs within each cell in terms of operational and structural information. The network-level similarity calculates OT distance between networks by considering both the cell-level similarity and the variation in the global position of each cell within their respective networks. We then explore HOTNN in a Bayesian optimization framework called HOTNAS, and demonstrate its efficacy in diverse tasks. Extensive experiments demonstrate that HOTNAS can discover network architectures with better performance in multiple modular cell-based search spaces.

We propose a method that leverages graph neural networks, multi-level message passing, and unsupervised training to enable real-time prediction of realistic clothing dynamics. Whereas existing methods based on linear blend skinning must be trained for specific garments, our method is agnostic to body shape and applies to tight-fitting garments as well as loose, free-flowing clothing. Our method furthermore handles changes in topology (e.g., garments with buttons or zippers) and material properties at inference time. As one key contribution, we propose a hierarchical message-passing scheme that efficiently propagates stiff stretching modes while preserving local detail. We empirically show that our method outperforms strong baselines quantitatively and that its results are perceived as more realistic than state-of-the-art methods.

Diffusion models have emerged as the best approach for generative modeling of 2D images. Part of their success is due to the possibility of training them on millions if not billions of images with a stable learning objective. However, extending these models to 3D remains difficult for two reasons. First, finding a large quantity of 3D training data is much more complex than for 2D images. Second, while it is conceptually trivial to extend the models to operate on 3D rather than 2D grids, the associated cubic growth in memory and compute complexity makes this infeasible. We address the first challenge by introducing a new diffusion setup that can be trained, end-to-end, with only posed 2D images for supervision; and the second challenge by proposing an image formation model that decouples model memory from spatial memory. We evaluate our method on real-world data, using the CO3D dataset which has not been used to train 3D generative models before. We show that our diffusion models are scalable, train robustly, and are competitive in terms of sample quality and fidelity to existing approaches for 3D generative modeling.

Human-Object Interaction (HOI) detection aims to localize human-object pairs and recognize their interactions. Recently, Contrastive Language-Image Pre-training (CLIP) has shown great potential in providing interaction prior for HOI detectors via knowledge distillation. However, such approaches often rely on large-scale training data and suffer from inferior performance under few/zero-shot scenarios. In this paper, we propose a novel HOI detection framework that efficiently extracts prior knowledge from CLIP and achieves better generalization. In detail, we first introduce a novel interaction decoder to extract informative regions in the visual feature map of CLIP via a cross-attention mechanism, which is then fused with the detection backbone by a knowledge integration block for more accurate human-object pair detection. In addition, prior knowledge in CLIP text encoder is leveraged to generate a classifier by embedding HOI descriptions. To distinguish fine-grained interactions, we build a verb classifier from training data via visual semantic arithmetic and a lightweight verb representation adapter. Furthermore, we propose a training-free enhancement to exploit global HOI predictions from CLIP. Extensive experiments demonstrate that our method outperforms the state of the art by a large margin on various settings, e.g. +4.04 mAP on HICO-Det. The source code is available in https://github.com/Artanic30/HOICLIP.

Implicit neural representations store videos as neural networks and have performed well for vision tasks such as video compression and denoising. With frame index and/or positional index as input, implicit representations (NeRV, E-NeRV, etc.) reconstruct video frames from fixed and content-agnostic embeddings. Such embedding largely limits the regression capacity and internal generalization for video interpolation. In this paper, we propose a Hybrid Neural Representation for Videos (HNeRV), where learnable and content-adaptive embeddings act as decoder input. Besides the input embedding, we introduce a HNeRV block to make model parameters evenly distributed across the entire network, therefore higher layers (layers near the output) can have more capacity to store high-resolution content and video details. With content-adaptive embedding and re-designed model architecture, HNeRV outperforms implicit methods (NeRV, E-NeRV) in video regression task for both reconstruction quality and convergence speed, and shows better internal generalization. As a simple and efficient video representation, HNeRV also shows decoding advantages for speed, flexibility, and deployment, compared to traditional codecs (H.264, H.265) and learning-based compression methods. Finally, we explore the effectiveness of HNeRV on downstream tasks such as video compression and video inpainting.

Supervision for metric learning has long been given in the form of equivalence between human-labeled classes. Although this type of supervision has been a basis of metric learning for decades, we argue that it hinders further advances in the field. In this regard, we propose a new regularization method, dubbed HIER, to discover the latent semantic hierarchy of training data, and to deploy the hierarchy to provide richer and more fine-grained supervision than inter-class separability induced by common metric learning losses. HIER achieves this goal with no annotation for the semantic hierarchy but by learning hierarchical proxies in hyperbolic spaces. The hierarchical proxies are learnable parameters, and each of them is trained to serve as an ancestor of a group of data or other proxies to approximate the semantic hierarchy among them. HIER deals with the proxies along with data in hyperbolic space since the geometric properties of the space are well-suited to represent their hierarchical structure. The efficacy of HIER is evaluated on four standard benchmarks, where it consistently improved the performance of conventional methods when integrated with them, and consequently achieved the best records, surpassing even the existing hyperbolic metric learning technique, in almost all settings.

Parsing an unstructured point set into constituent local geometry structures (e.g., edges or surfaces) would be helpful for understanding and representing point clouds. This motivates us to design a deep architecture to model the hierarchical geometry from points, edges, surfaces (triangles), to super-surfaces (adjacent surfaces) for the thorough analysis of point clouds. In this paper, we present a novel Hierarchical Geometry Network (HGNet) that integrates such hierarchical geometry structures from super-surfaces, surfaces, edges, to points in a top-down manner for learning point cloud representations. Technically, we first construct the edges between every two neighbor points. A point-level representation is learnt with edge-to-point aggregation, i.e., aggregating all connected edges into the anchor point. Next, as every two neighbor edges compose a surface, we obtain the edge-level representation of each anchor edge via surface-to-edge aggregation over all neighbor surfaces. Furthermore, the surface-level representation is achieved through super-surface-to-surface aggregation by transforming all super-surfaces into the anchor surface. A Transformer structure is finally devised to unify all the point-level, edge-level, and surface-level features into the holistic point cloud representations. Extensive experiments on four point cloud analysis datasets demonstrate the superiority of HGNet for 3D object classification and part/semantic segmentation tasks. More remarkably, HGNet achieves the overall accuracy of 89.2% on ScanObjectNN, improving PointNeXt-S by 1.5%.

Current semantic segmentation models have achieved great success under the independent and identically distributed (i.i.d.) condition. However, in real-world applications, test data might come from a different domain than training data. Therefore, it is important to improve model robustness against domain differences. This work studies semantic segmentation under the domain generalization setting, where a model is trained only on the source domain and tested on the unseen target domain. Existing works show that Vision Transformers are more robust than CNNs and show that this is related to the visual grouping property of self-attention. In this work, we propose a novel hierarchical grouping transformer (HGFormer) to explicitly group pixels to form part-level masks and then whole-level masks. The masks at different scales aim to segment out both parts and a whole of classes. HGFormer combines mask classification results at both scales for class label prediction. We assemble multiple interesting cross-domain settings by using seven public semantic segmentation datasets. Experiments show that HGFormer yields more robust semantic segmentation results than per-pixel classification methods and flat-grouping transformers, and outperforms previous methods significantly. Code will be available at https://github.com/dingjiansw101/HGFormer.

While today's high dynamic range (HDR) image fusion algorithms are capable of blending multiple exposures, the acquisition is often controlled so that the dynamic range within one exposure is narrow. For HDR imaging in photon-limited situations, the dynamic range can be enormous and the noise within one exposure is spatially varying. Existing image denoising algorithms and HDR fusion algorithms both fail to handle this situation, leading to severe limitations in low-light HDR imaging. This paper presents two contributions. Firstly, we identify the source of the problem. We find that the issue is associated with the co-existence of (1) spatially varying signal-to-noise ratio, especially the excessive noise due to very dark regions, and (2) a wide luminance range within each exposure. We show that while the issue can be handled by a bank of denoisers, the complexity is high. Secondly, we propose a new method called the spatially varying high dynamic range (SV-HDR) fusion network to simultaneously denoise and fuse images. We introduce a new exposure-shared block within our custom-designed multi-scale transformer framework. In a variety of testing conditions, the performance of the proposed SV-HDR is better than the existing methods.

We present HARP (HAnd Reconstruction and Personalization), a personalized hand avatar creation approach that takes a short monocular RGB video of a human hand as input and reconstructs a faithful hand avatar exhibiting a high-fidelity appearance and geometry. In contrast to the major trend of neural implicit representations, HARP models a hand with a mesh-based parametric hand model, a vertex displacement map, a normal map, and an albedo without any neural components. The explicit nature of our representation enables a truly scalable, robust, and efficient approach to hand avatar creation as validated by our experiments. HARP is optimized via gradient descent from a short sequence captured by a hand-held mobile phone and can be directly used in AR/VR applications with real-time rendering capability. To enable this, we carefully design and implement a shadow-aware differentiable rendering scheme that is robust to high degree articulations and self-shadowing regularly present in hand motions, as well as challenging lighting conditions. It also generalizes to unseen poses and novel viewpoints, producing photo-realistic renderings of hand animations. Furthermore, the learned HARP representation can be used for improving 3D hand pose estimation quality in challenging viewpoints. The key advantages of HARP are validated by the in-depth analyses on appearance reconstruction, novel view and novel pose synthesis, and 3D hand pose refinement. It is an AR/VR-ready personalized hand representation that shows superior fidelity and scalability.

A great deal of progress has been made in image captioning, driven by research into how to encode the image using pre-trained models. This includes visual encodings (e.g. image grid features or detected objects) and more recently textual encodings (e.g. image tags or text descriptions of image regions). As more advanced encodings are available and incorporated, it is natural to ask: how to efficiently and effectively leverage the heterogeneous set of encodings? In this paper, we propose to regard the encodings as augmented views of the input image. The image captioning model encodes each view independently with a shared encoder efficiently, and a contrastive loss is incorporated across the encoded views in a novel way to improve their representation quality and the model's data efficiency. Our proposed hierarchical decoder then adaptively weighs the encoded views according to their effectiveness for caption generation by first aggregating within each view at the token level, and then across views at the view level. We demonstrate significant performance improvements of +5.6% CIDEr on MS-COCO and +12.9% CIDEr on Flickr30k compared to state of the arts,

Real-time 3D hand mesh reconstruction is challenging, especially when the hand is holding some object. Beyond the previous methods, we design H2ONet to fully exploit non-occluded information from multiple frames to boost the reconstruction quality. First, we decouple hand mesh reconstruction into two branches, one to exploit finger-level non-occluded information and the other to exploit global hand orientation, with lightweight structures to promote real-time inference. Second, we propose finger-level occlusion-aware feature fusion, leveraging predicted finger-level occlusion information as guidance to fuse finger-level information across time frames. Further, we design hand-level occlusion-aware feature fusion to fetch non-occluded information from nearby time frames. We conduct experiments on the Dex-YCB and HO3D-v2 datasets with challenging hand-object occlusion cases, manifesting that H2ONet is able to run in real-time and achieves state-of-the-art performance on both the hand mesh and pose precision. The code will be released on GitHub.

Standard Unsupervised Domain Adaptation (UDA) methods assume the availability of both source and target data during the adaptation. In this work, we investigate Source-free Unsupervised Domain Adaptation (SF-UDA), a specific case of UDA where a model is adapted to a target domain without access to source data. We propose a novel approach for the SF-UDA setting based on a loss reweighting strategy that brings robustness against the noise that inevitably affects the pseudo-labels. The classification loss is reweighted based on the reliability of the pseudo-labels that is measured by estimating their uncertainty. Guided by such reweighting strategy, the pseudo-labels are progressively refined by aggregating knowledge from neighbouring samples. Furthermore, a self-supervised contrastive framework is leveraged as a target space regulariser to enhance such knowledge aggregation. A novel negative pairs exclusion strategy is proposed to identify and exclude negative pairs made of samples sharing the same class, even in presence of some noise in the pseudo-labels. Our method outperforms previous methods on three major benchmarks by a large margin. We set the new SF-UDA state-of-the-art on VisDA-C and DomainNet with a performance gain of +1.8% on both benchmarks and on PACS with +12.3% in the single-source setting and +6.6% in multi-target adaptation. Additional analyses demonstrate that the proposed approach is robust to the noise, which results in significantly more accurate pseudo-labels compared to state-of-the-art approaches.

Model selection is essential for reducing the search cost of the best pre-trained model over a large-scale model zoo for a downstream task. After analyzing recent hand-designed model selection criteria with 400+ ImageNet pre-trained models and 40 downstream tasks, we find that they can fail due to invalid assumptions and intrinsic limitations. The prior knowledge on model capacity and dataset also can not be easily integrated into the existing criteria. To address these issues, we propose to convert model selection as a recommendation problem and to learn from the past training history. Specifically, we characterize the meta information of datasets and models as features, and use their transfer learning performance as the guided score. With thousands of historical training jobs, a recommendation system can be learned to predict the model selection score given the features of the dataset and the model as input. Our approach enables integrating existing model selection scores as additional features and scales with more historical data. We evaluate the prediction accuracy with 22 pre-trained models over 40 downstream tasks. With extensive evaluations, we show that the learned approach can outperform prior hand-designed model selection methods significantly when relevant training history is available.