ICCV 2021 Papers

Point cloud completion aims to predict a complete shape in high accuracy from its partial observation. However, previous methods usually suffered from discrete nature of point cloud and unstructured prediction of points in local regions, which makes it hard to reveal fine local geometric details on the complete shape. To resolve this issue, we propose SnowflakeNet with Snowflake Point Deconvolution (SPD) to generate the complete point clouds. The SnowflakeNet models the generation of complete point clouds as the snowflake-like growth of points in 3D space, where the child points are progressively generated by splitting their parent points after each SPD. Our insight of revealing detailed geometry is to introduce skip-transformer in SPD to learn point splitting patterns which can fit local regions the best. Skip-transformer leverages attention mechanism to summarize the splitting patterns used in the previous SPD layer to produce the splitting in the current SPD layer. The locally compact and structured point cloud generated by SPD is able to precisely capture the structure characteristic of 3D shape in local patches, which enables the network to predict highly detailed geometries, such as smooth regions, sharp edges and corners. Our experimental results outperform the state-of-the-art point cloud completion methods under widely used benchmarks. Code will be available at https://github.com/AllenXiangX/SnowflakeNet.

SmartShadow is a deep learning application for digital painting artists to draw shadows on line drawings, with three proposed tools. (1) Shadow brush: artists can draw scribbles to coarsely indicate the areas inside or outside their wanted shadows, and the application will generate the shadows in real-time. (2) Shadow boundary brush: this brush can precisely control the boundary of any specific shadow. (3) Global shadow generator: this tool can estimate the global shadow direction from input brush scribbles, and then consistently propagate local shadows to the entire image. These three tools can not only speed up the shadow drawing process (by 3.1 times as experiments validate), but also allow for the flexibility to achieve various shadow effects and facilitate richer artistic creations. To this end, we train Convolutional Neural Networks (CNNs) with a collected large-scale dataset of both real and synthesized data, and especially, we collect 1670 shadow samples drawn by real artists. Both qualitative analysis and user study show that our approach can generate high-quality shadows that are practically usable in the daily works of digital painting artists. We present 30 additional results and 15 visual comparisons in the supplementary materiel.

The key challenge in designing a sketch representation lies with handling the abstract and iconic nature of sketches. Existing work predominantly utilizes either, (i) a pixelative format that treats sketches as natural images employing off-the-shelf CNN-based networks, or (ii) an elaborately designed vector format that leverages the structural information of drawing orders using sequential RNN-based methods. While the pixelative format lacks intuitive exploitation of structural cues, sketches in vector format are absent in most cases limiting their practical usage. Hence, in this paper, we propose a lattice structured sketch representation that not only removes the bottleneck of requiring vector data but also preserves the structural cues that vector data provides. Essentially, sketch lattice is a set of points sampled from the pixelative format of the sketch using a lattice graph. We show that our lattice structure is particularly amenable to structural changes that largely benefits sketch abstraction modeling for generation tasks. Our lattice representation could be effectively encoded using a graph model, that uses significantly fewer model parameters (13.5 times lesser) than existing state-of-the-art. Extensive experiments demonstrate the effectiveness of sketch lattice for sketch manipulation, including sketch healing and image-to-sketch synthesis.

What makes free-hand sketches appealing for humans lies with its capability as a universal tool to depict the visual world. Such flexibility at human ease, however, introduces abstract renderings that pose unique challenges to computer vision models. In this paper, we propose a purpose-made sketch representation for human sketches. The key intuition is that such representation should be abstract at design, so to accommodate the abstract nature of sketches. This is achieved by interpreting sketch abstraction on two levels: appearance and structure. We abstract sketch structure as a pre-defined coarse-to-fine visual block hierarchy, and average visual features within each block to model appearance abstraction. We then discuss three general strategies on how to exploit feature synergy across different levels of this abstraction hierarchy. The superiority of explicitly abstracting sketch representation is empirically validated on a number of sketch analysis tasks, including sketch recognition, fine-grained sketch-based image retrieval, and generative sketch healing. Our simple design not only yields strong results on all said tasks, but also offers intuitive feature granularity control to tailor for various downstream tasks. Code will be made publicly available.

Reconstructing 3D shape from 2D sketches has long been an open problem because the sketches only provide very sparse and ambiguous information. In this paper, we use an encoder/decoder architecture for the sketch to mesh translation. When integrated into a user interface that provides camera parameters for the sketches, this enables us to leverage its latent parametrization to represent and refine a 3D mesh so that its projections match the external contours outlined in the sketch. We will show that this approach is easy to deploy, robust to style changes, and effective. Furthermore, it can be used for shape refinement given only single pen strokes. We compare our approach to state-of-the-art methods on sketches - both hand-drawn and synthesized - and demonstrate that we outperform them.

Can a user create a deep generative model by sketching a single example? Traditionally, creating a GAN model has required the collection of a large-scale dataset of exemplars and specialized knowledge in deep learning. In contrast, sketching is possibly the most universally accessible way to convey a visual concept. In this work, we present a method, GAN Sketching, for rewriting GANs with one or more sketches, to make GANs training easier for novice users. In particular, we change the weights of an original GAN model according to user sketches. We encourage the model's output to match the user sketches through a cross-domain adversarial loss. Furthermore, we explore different regularization methods to preserve the original model's diversity and image quality. Experiments have shown that our method can mold GANs to match shapes and poses specified by sketches while maintaining realism and diversity. Finally, we demonstrate a few applications of the resulting GAN, including latent space interpolation and image editing.

In this paper, we decouple unsupervised human mesh recovery into the well-studied problems of unsupervised 3D pose estimation, and human mesh recovery from estimated 3D skeletons, focusing on the latter task. The challenges of the latter task are two folds: (1) pose failure (i.e., pose mismatching -- different skeleton definitions in dataset and SMPL , and pose ambiguity -- endpoints have arbitrary joint angle configurations for the same 3D joint coordinates). (2) shape ambiguity (i.e., the lack of shape constraints on body configuration). To address these issues, we propose Skeleton2Mesh, a novel lightweight framework that recovers human mesh from a single image. Our Skeleton2Mesh contains three modules, i.e., Differentiable Inverse Kinematics (DIK), Pose Refinement (PR) and Shape Refinement (SR) modules. DIK is designed to transfer 3D rotation from estimated 3D skeletons, which relies on a minimal set of kinematics prior knowledge. Then PR and SR modules are utilized to tackle the pose ambiguity and shape ambiguity respectively. All three modules can be incorporated into Skeleton2Mesh seamlessly via an end-to-end manner. Furthermore, we utilize an adaptive joint regressor to alleviate the effects of skeletal topology from different datasets. Results on the Human3.6M dataset for human mesh recovery demonstrate that our method improves upon the previous unsupervised methods by 32.6% under the same setting. Qualitative results on in-the-wild datasets exhibit that the recovered 3D meshes are natural, realistic. Our project is available at https://sites.google.com/view/skeleton2mesh.

Skeleton-based human action recognition has attracted increasing attention in recent years. However, most of the existing works focus on supervised learning which requiring a large number of annotated action sequences that are often expensive to collect. We investigate unsupervised representation learning for skeleton action recognition, and design a novel skeleton cloud colorization technique that is capable of learning skeleton representations from unlabeled skeleton sequence data. Specifically, we represent a skeleton action sequence as a 3D skeleton cloud and colorize each point in the cloud according to its temporal and spatial orders in the original (unannotated) skeleton sequence. Leveraging the colorized skeleton point cloud, we design an auto-encoder framework that can learn spatial-temporal features from the artificial color labels of skeleton joints effectively. We evaluate our skeleton cloud colorization approach with action classifiers trained under different configurations, including unsupervised, semi-supervised and fully-supervised settings. Extensive experiments on NTU RGB+D and NW-UCLA datasets show that the proposed method outperforms existing unsupervised and semi-supervised 3D action recognition methods by large margins, and it achieves competitive performance in supervised 3D action recognition as well.

Imaging depth and spectrum have been extensively studied in isolation from each other for decades. Recently, hyperspectral-depth (HS-D) imaging emerges to capture both information simultaneously by combining two different imaging systems; one for depth, the other for spectrum. While being accurate, this combinational approach induces increased form factor, cost, capture time, and alignment/registration problems. In this work, departing from the combinational principle, we propose a compact single-shot monocular HS-D imaging method. Our method uses a diffractive optical element (DOE), the point spread function of which changes with respect to both depth and spectrum. This enables us to reconstruct spectrum and depth from a single captured image. To this end, we develop a differentiable simulator and a neural-network-based reconstruction method that are jointly optimized via automatic differentiation. To facilitate learning the DOE, we present a first HS-D dataset by building a benchtop HS-D imager that acquires high-quality ground truth. We evaluate our method with synthetic and real experiments by building an experimental prototype and achieve state-of-the-art HS-D imaging results.

We propose a fully automated system that simultaneously estimates the camera intrinsics, the ground plane, and physical distances between people from a single RGB image or video captured by a camera viewing a 3-D scene from a fixed vantage point. To automate camera calibration and distance estimation, we leverage priors about human pose and develop a novel direct formulation for pose-based auto-calibration and distance estimation, which shows state-of-the-art performance on publicly available datasets. The proposed approach enables existing camera systems to measure physical distances without needing a dedicated calibration process or range sensors, and is applicable to a broad range of use cases such as social distancing and workplace safety. Furthermore, to enable evaluation and drive research in this area, we contribute to the publicly available MEVA dataset with additional distance annotations, resulting in "MEVADA" -- an evaluation benchmark for the pose-based auto-calibration and distance estimation problem.

This paper proposes a novel deep learning approach for single image defocus deblurring based on inverse kernels. In a defocused image, the blur shapes are similar among pixels although the blur sizes can spatially vary. To utilize the property with inverse kernels, we exploit the observation that when only the size of a defocus blur changes while keeping the shape, the shape of the corresponding inverse kernel remains the same and only the scale changes. Based on the observation, we propose a kernel-sharing parallel atrous convolutional (KPAC) block specifically designed by incorporating the property of inverse kernels for single image defocus deblurring. To effectively simulate the invariant shapes of inverse kernels with different scales, KPAC shares the same convolutional weights among multiple atrous convolution layers. To efficiently simulate the varying scales of inverse kernels, KPAC consists of only a few atrous convolution layers with different dilations and learns per-pixel scale attentions to aggregate the outputs of the layers. KPAC also utilizes the shape attention to combine the outputs of multiple convolution filters in each atrous convolution layer, to deal with defocus blur with a slightly varying shape. We demonstrate that our approach achieves state-of-the-art performance with a much smaller number of parameters than previous methods.

In this work, we tackle the problem of single image-based 3D shape retrieval (IBSR), where we seek to find the most matched shape of a given single 2D image from a shape repository. Most of the existing works learn to embed 2D images and 3D shapes into a common feature space and perform metric learning using a triplet loss. Inspired by the great success in recent contrastive learning works on self-supervised representation learning, we propose a novel IBSR pipeline leveraging contrastive learning. We note that adopting such cross-modal contrastive learning between 2D images and 3D shapes into IBSR tasks is non-trivial and challenging: contrastive learning requires very strong data augmentation in constructed positive pairs to learn the feature invariance, whereas traditional metric learning works do not have this requirement. Moreover, object shape and appearance are entangled in 2D query images, thus making the learning task more difficult than contrasting single-modal data. To mitigate the challenges, we propose to use multi-view grayscale rendered images from the 3D shapes as a shape representation. We then introduce a strong data augmentation technique based on color transfer, which can significantly but naturally change the appearance of the query image, effectively satisfying the need for contrastive learning. Finally, we propose to incorporate a novel category-level contrastive loss that helps distinguish similar objects from different categories, in addition to classic instance-level contrastive loss. Our experiments demonstrate that our approach achieves the best performance on all the three popular IBSR benchmarks, including Pix3D, Stanford Cars, and Comp Cars, outperforming the previous state-of-the-art from 4% - 15% on retrieval accuracy.

A few-shot semantic segmentation model is typically composed of a CNN encoder, a CNN decoder and a simple classifier (separating foreground and background pixels). Most existing methods meta-learn all three model components for fast adaptation to a new class. However, given that as few as a single support set image is available, effective model adaption of all three components to the new class is extremely challenging. In this work we propose to simplify the meta-learning task by focusing solely on the simplest component -- the classifier, whilst leaving the encoder and decoder to pre-training. We hypothesize that if we pre-train an off-the-shelf segmentation model over a set of diverse training classes with sufficient annotations, the encoder and decoder can capture rich discriminative features applicable for any unseen classes, rendering the subsequent meta-learning stage unnecessary. For the classifier meta-learning, we introduce a Classifier Weight Transformer (CWT) designed to dynamically adapt the support-set trained classifier's weights to each query image in an inductive way. Extensive experiments on two standard benchmarks show that despite its simplicity, our method outperforms the state-of-the-art alternatives, often by a large margin. Code is available on https://github.com/zhiheLu/CWT-for-FSS.

This paper presents a Simple and effective unsupervised adaptation method for Robust Object Detection (SimROD). To overcome the challenging issues of domain shift and pseudo-label noise, our method integrates a novel domain-centric data augmentation, a gradual self-labeling adaptation procedure, and a teacher-guided fine-tuning mechanism. Using our method, target domain samples can be leveraged to adapt object detection models without changing the model architecture or generating synthetic data. When applied to image corruptions and high-level cross-domain adaptation benchmarks, our method outperforms prior baselines on multiple domain adaptation benchmarks. SimROD achieves new state-of-the-art on standard real-to-synthetic and cross-camera setup benchmarks. On the image corruption benchmark, models adapted with our method achieved a relative robustness improvement of 15-25% AP50 on Pascal-C and 5-6% AP on COCO-C and Cityscapes-C. On the cross-domain benchmark, our method outperformed the best baseline performance by up to 8% and 4% AP50 on Comic and Watercolor respectively.

Hand gesture serves as a critical role in sign language. Current deep-learning-based sign language recognition (SLR) methods may suffer insufficient interpretability and overfitting due to limited sign data sources. In this paper, we introduce the first self-supervised pre-trainable SignBERT with incorporated hand prior for SLR. SignBERT views the hand pose as a visual token, which is derived from an off-the-shelf pose extractor. The visual tokens are then embedded with gesture state, temporal and hand chirality information. To take full advantage of available sign data sources, SignBERT first performs self-supervised pre-training by masking and reconstructing visual tokens. Jointly with several mask modeling strategies, we attempt to incorporate hand prior in a model-aware method to better model hierarchical context over the hand sequence. Then with the prediction head added, SignBERT is fine-tuned to perform the downstream SLR task. To validate the effectiveness of our method on SLR, we perform extensive experiments on four public benchmark datasets, i.e., NMFs-CSL, SLR500, MSASL and WLASL. Experiment results demonstrate the effectiveness of both self-supervised learning and imported hand prior. Furthermore, we achieve state-of-the-art performance on all benchmarks with a notable gain.

RGB-D semantic segmentation has attracted increasing attention over the past few years. Existing methods mostly employ homogeneous convolution operators to consume the RGB and depth features, ignoring their intrinsic differences. In fact, the RGB values capture the photometric appearance properties in the projected image space, while the depth feature encodes both the shape of a local geometry as well as the base (whereabout) of it in a larger context. Compared with the base, the shape probably is more inherent and has a stronger connection to the semantics, and thus is more critical for segmentation accuracy. Inspired by this observation, we introduce Shape-aware Convolutional layer (ShapeConv) for processing the depth feature, where the depth feature is firstly decomposed into a shape-component and a base-component, next two learnable weights are introduced to cooperate with them independently, and finally a convolution is applied on the re-weighted combination of these two components. ShapeConv is model-agnostic and can be easily integrated into most CNNs to replace vanilla convolutional layers for semantic segmentation. Extensive experiments on three challenging indoor RGB-D semantic segmentation benchmarks, i.e., NYU-Dv2(-13,-40), SUN RGB-D, and SID, demonstrate the effectiveness of our ShapeConv when employing it over five popular architectures. Moreover, the performance of CNNs with ShapeConv is boosted without introducing any computation and memory increase in the inference phase. The reason is that the learnt weights for balancing the importance between the shape and base components in ShapeConv become constants in the inference phase, and thus can be fused into the following convolution, resulting in a network that is identical to one with vanilla convolutional layers.

There is an emerging sense that the vulnerability of Image Convolutional Neural Networks (CNN), i.e., sensitivity to image corruptions, perturbations, and adversarial attacks, is connected with Texture Bias. This relative lack of Shape Bias is also responsible for poor performance in Domain Generalization (DG). The inclusion of a role of shape alleviates these vulnerabilities and some approaches have achieved this by training on negative images, images endowed with edge maps, or images with conflicting shape and texture information. This paper advocates an explicit and complete representation of shape using a classical computer vision approach, namely, representing the shape content of an image with the shock graph of its contour map. The resulting graph and its descriptor is a complete representation of contour content and is classified using recent Graph Neural Network (GNN) methods. The experimental results on three domain shift datasets, Colored MNIST, PACS, and VLCS demonstrate that even without using appearance the shape-based approach exceeds classical Image CNN based methods in domain generalization.

In this paper we contribute a simple yet effective approach for estimating 3D poses of multiple people from multi-view images. Our proposed coarse-to-fine pipeline first aggregates noisy 2D observations from multiple camera views into 3D space and then associates them into individual instances based on a confidence-aware majority voting technique. The final pose estimates are attained from a novel optimization scheme which links high-confidence multi-view 2D observations and 3D joint candidates. Moreover, a statistical parametric body model such as SMPL is leveraged as a regularizing prior for these 3D joint candidates. Specifically, both 3D poses and SMPL parameters are optimized jointly in an alternating fashion. Here the parametric models help in correcting implausible 3D pose estimates and filling in missing joint detections while updated 3D poses in turn guide obtaining better SMPL estimations. By linking 2D and 3D observations, our method is both accurate and generalizes to different data sources because it better decouples the final 3D pose from the inter-person constellation and is more robust to noisy 2D detections. We systematically evaluate our method on public datasets and achieve state-of-the-art performance. The code and video will be available on the project page: https://ait.ethz.ch/projects/2021/multi-human-pose/.

We develop a novel self-supervised learning method named Shape Self-Correction for point cloud analysis. Our method is motivated by the principle that a good shape representation should be able to find distorted parts of a shape and correct them. To learn strong shape representations in an unsupervised manner, we first design a shape-disorganizing module to destroy certain local shape parts of an object. Then the destroyed shape and the normal shape are sent into a point cloud network to get representations, which are employed to segment points that belong to distorted parts and further reconstruct them to restore the shape to normal. To perform better in these two associated pretext tasks, the network is constrained to capture useful shape features from the object, which indicates that the point cloud network encodes rich geometric and contextual information. The learned feature extractor transfers well to downstream classification and segmentation tasks. Experimental results on ModelNet, ScanNet and ShapeNetPart demonstrate that our method achieves state-of-the-art performance among unsupervised methods. Our framework can be applied to a wide range of deep learning networks for point cloud analysis and we show experimentally that pre-training with our framework significantly boosts the performance of supervised models.