CVPR 2022 Papers

Reconstructing the 3D pose of a person in metric scale from a single view image is a geometrically ill-posed problem. For example, we can not measure the exact distance of a person to the camera from a single view image without additional scene assumptions (e.g., known height). Existing learning based approaches circumvent this issue by reconstructing the 3D pose up to scale. However, there are many applications such as virtual telepresence, robotics, and augmented reality that require metric scale reconstruction. In this paper, we show that audio signals recorded along with an image, provide complementary information to reconstruct the metric 3D pose of the person. The key insight is that as the audio signals traverse across the 3D space, their interactions with the body provide metric information about the body's pose. Based on this insight, we introduce a time-invariant transfer function called pose kernel---the impulse response of audio signals induced by the body pose. The main properties of the pose kernel are that (1) its envelope highly correlates with 3D pose, (2) the time response corresponds to arrival time, indicating the metric distance to the microphone, and (3) it is invariant to changes in the scene geometry configurations. Therefore, it is readily generalizable to unseen scenes. We design a multi-stage 3D CNN that fuses audio and visual signals and learns to reconstruct 3D pose in a metric scale. We show that our multi-modal method produces accurate metric reconstruction in real world scenes, which is not possible with state-of-the-art lifting approaches including parametric mesh regression and depth regression.

In portraits, eyeglasses may occlude facial regions and generate cast shadows on faces, which degrades the performance of many techniques like face verification and expression recognition. Portrait eyeglasses removal is critical in handling these problems. However, completely removing the eyeglasses is challenging because the lighting effects (e.g., cast shadows) caused by them are often complex. In this paper, we propose a novel framework to remove eyeglasses as well as their cast shadows from face images. The method works in a detect-then-remove manner, in which eyeglasses and cast shadows are both detected and then removed from images. Due to the lack of paired data for supervised training, we present a new synthetic portrait dataset with both intermediate and final supervisions for both the detection and removal tasks. Furthermore, we apply a cross-domain technique to fill the gap between the synthetic and real data. To the best of our knowledge, the proposed technique is the first to remove eyeglasses and their cast shadows simultaneously. The code and synthetic dataset are available at https://github.com/StoryMY/take-off-eyeglasses.

We propose a framework that can deform an object in a 2D image as it exists in 3D space. Most existing methods for 3D-aware image manipulation are limited to (1) only changing the global scene information or depth, or (2) manipulating an object of specific categories. In this paper, we present a 3D-aware image deformation method with minimal restrictions on shape category and deformation type. While our framework leverages 2D-to-3D reconstruction, we argue that reconstruction is not sufficient for realistic deformations due to the vulnerability to topological errors. Thus, we propose to take a supervised learning-based approach to predict the shape Laplacian of the underlying volume of a 3D reconstruction represented as a point cloud. Given the deformation energy calculated using the predicted shape Laplacian and user-defined deformation handles (e.g., keypoints), we obtain bounded biharmonic weights to model plausible handle-based image deformation. In the experiments, we present our results of deforming 2D character and clothed human images. We also quanti- tatively show that our approach can produce more accurate deformation weights compared to alternative methods (i.e., mesh reconstruction and point cloud Laplacian methods).

The size and shape of the receptive field determine how the network aggregates local features, and affect the overall performance of a model considerably. Many components in a neural network, such as depth, kernel sizes, and strides for convolution and pooling, influence the receptive field. However, they still rely on hyperparameters, and the receptive fields of existing models result in suboptimal shapes and sizes. Hence, we propose a simple yet effective Dynamically Optimized Pooling operation, referred to as DynOPool, which learns the optimized scale factors of feature maps end-to-end. Moreover, DynOPool determines the proper resolution of a feature map by learning the desirable size and shape of its receptive field, which allows an operator in a deeper layer to observe an input image in the optimal scale. Any kind of resizing modules in a deep neural network can be replaced by DynOPool with minimal cost. Also, DynOPool controls the complexity of the model by introducing an additional loss term that constrains computational cost. Our experiments show that the models equipped with the proposed learnable resizing module outperform the baseline algorithms on multiple datasets in image classification and semantic segmentation.

Modern image generative models show remarkable sample quality when trained on a single domain or class of objects. In this work, we introduce a generative adversarial network that can simultaneously generate aligned image samples from multiple related domains. We leverage the fact that a variety of object classes share common attributes, with certain geometric differences. We propose Polymorphic-GAN which learns shared features across all domains and a per-domain morph layer to morph shared features according to each domain. In contrast to previous works, our framework allows simultaneous modelling of images with highly varying geometries, such as images of human faces, painted and artistic faces, as well as multiple different animal faces. We demonstrate that our model produces aligned samples for all domains and show how it can be used for applications such as segmentation transfer and cross-domain image editing, as well as training in low-data regimes. Additionally, we apply our Polymorphic-GAN on image-to-image translation tasks and show that we can greatly surpass previous approaches in cases where the geometric differences between domains are large.

While most state-of-the-art instance segmentation methods produce binary segmentation masks, geographic and cartographic applications typically require precise vector polygons of extracted objects instead of rasterized output. This paper introduces PolyWorld, a neural network that directly extracts building vertices from an image and connects them correctly to create precise polygons. The model predicts the connection strength between each pair of vertices using a graph neural network and estimates the assignments by solving a differentiable optimal transport problem. Moreover, the vertex positions are optimized by minimizing a combined segmentation and polygonal angle difference loss. PolyWorld significantly outperforms the state of the art in building polygonization and achieves not only notable quantitative results, but also produces visually pleasing building polygons. Code and trained weights are publicly available at https://github.com/zorzi-s/PolyWorldPretrainedNetwork.

We present Polarity Sampling, a theoretically justified plug-and-play method for controlling the generation quality and diversity of any pre-trained deep generative network (DGN). Leveraging the fact that DGNs are, or can be approximated by, continuous piecewise affine splines, we derive the analytical DGN output space distribution as a function of the product of the DGN's Jacobian singular values raised to a power rho. We dub rho the polarity parameter and prove that rho focuses the DGN sampling on the modes (rho < 0) or anti-modes (rho > 0) of the DGN output space probability distribution. We demonstrate that nonzero polarity values achieve a better precision-recall (quality-diversity) Pareto frontier than standard methods, such as truncation, for a number of state-of-the-art DGNs. We also present quantitative and qualitative results on the improvement of overall generation quality (e.g., in terms of the Frechet Inception Distance) for a number of state-of-the-art DGNs, including StyleGAN3, BigGAN-deep, NVAE, for different conditional and unconditional image generation tasks. In particular, Polarity Sampling redefines the state-of-the-art for StyleGAN2 on the FFHQ Dataset to FID 2.57, StyleGAN2 on the LSUN Car Dataset to FID 2.27 and StyleGAN3 on the AFHQv2 Dataset to FID 3.95. Colab Demo: bit.ly/polarity-samp

Optimization of Top-1 ImageNet promotes enormous networks that may be impractical in inference settings. Binary neural networks (BNNs) have the potential to significantly lower the compute intensity but existing models suffer from low quality. To overcome this deficiency, we propose PokeConv, a binary convolution block which improves quality of BNNs by techniques such as adding multiple residual paths, and tuning the activation function. We apply it to ResNet-50 and optimize ResNet's initial convolutional layer which is hard to binarize. We name the resulting network family PokeBNN. These techniques are chosen to yield favorable improvements in both top-1 accuracy and the network's cost. In order to enable joint optimization of the cost together with accuracy, we define arithmetic computation effort (ACE), a hardware- and energy-inspired cost metric for quantized and binarized networks. We also identify a need to optimize an under-explored hyper-parameter controlling the binarization gradient approximation. We establish a new, strong state-of-the-art (SOTA) on top-1 accuracy together with commonly-used CPU64 cost, ACE cost and network size metrics. ReActNet-Adam, the previous SOTA in BNNs, achieved a 70.5% top-1 accuracy with 7.9 ACE. A small variant of PokeBNN achieves 70.5% top-1 with 2.6 ACE, more than 3x reduction in cost; a larger PokeBNN achieves 75.6% top-1 with 7.8 ACE, more than 5% improvement in accuracy without increasing the cost. PokeBNN implementation in JAX / Flax and reproduction instructions are open sourced. Source code and reproduction instructions are available in AQT repository: https://github.com/google/aqt.

We propose an embarrassingly simple point annotation scheme to collect weak supervision for instance segmentation. In addition to bounding boxes, we collect binary labels for a set of points uniformly sampled inside each bounding box. We show that the existing instance segmentation models developed for full mask supervision can be seamlessly trained with point-based supervision collected via our scheme. Remarkably, Mask R-CNN trained on COCO, PASCAL VOC, Cityscapes, and LVIS with only 10 annotated random points per object achieves 94%-98% of its fully-supervised performance, setting a strong baseline for weakly-supervised instance segmentation. The new point annotation scheme is approximately 5 times faster than annotating full object masks, making high-quality instance segmentation more accessible in practice. Inspired by the point-based annotation form, we propose a modification to PointRend instance segmentation module. For each object, the new architecture, called Implicit PointRend, generates parameters for a function that makes the final point-level mask prediction. Implicit PointRend is more straightforward and uses a single point-level mask loss. Our experiments show that the new module is more suitable for the point-based supervision.

Recently, zero-shot and few-shot learning via Contrastive Vision-Language Pre-training (CLIP) have shown inspirational performance on 2D visual recognition, which learns to match images with their corresponding texts in open-vocabulary settings. However, it remains under explored that whether CLIP, pre-trained by large-scale image-text pairs in 2D, can be generalized to 3D recognition. In this paper, we identify such a setting is feasible by proposing PointCLIP, which conducts alignment between CLIP-encoded point cloud and 3D category texts. Specifically, we encode a point cloud by projecting it into multi-view depth maps without rendering, and aggregate the view-wise zero-shot prediction to achieve knowledge transfer from 2D to 3D. On top of that, we design an inter-view adapter to better extract the global feature and adaptively fuse the few-shot knowledge learned from 3D into CLIP pre-trained in 2D. By just fine-tuning the lightweight adapter in the few-shot settings, the performance of PointCLIP could be largely improved. In addition, we observe the complementary property between PointCLIP and classical 3D-supervised networks. By simple ensembling, PointCLIP boosts baseline's performance and even surpasses state-of-the-art models. Therefore, PointCLIP is a promising alternative for effective 3D point cloud understanding via CLIP under low resource cost and data regime. We conduct thorough experiments on widely-adopted ModelNet10, ModelNet40 and the challenging ScanObjectNN to demonstrate the effectiveness of PointCLIP.

We present a simple and effective framework, named Point2Seq, for 3D object detection from point clouds. In contrast to previous methods that normally predict attributes of 3D objects all at once, we expressively model the interdependencies between attributes of 3D objects, which in turn enables a better detection accuracy. Specifically, we view each 3D object as a sequence of words and reformulate the 3D object detection task as decoding words from 3D scenes in an auto-regressive manner. We further propose a lightweight scene-to-sequence decoder that can auto-regressively generate words conditioned on features from a 3D scene as well as cues from the preceding words. The predicted words eventually constitute a set of sequences that completely describe the 3D objects in the scene, and all the predicted sequences are then automatically assigned to the respective ground truths through similarity-based sequence matching. Our approach is conceptually intuitive and can be readily plugged upon most existing 3D-detection backbones without adding too much computational overhead; the sequential decoding paradigm we proposed, on the other hand, can better exploit information from complex 3D scenes with the aid of preceding predicted words. Without bells and whistles, our method significantly outperforms the previous anchor- and center-based 3D object detection frameworks, yielding the new state-of-the-art on the challenging ONCE dataset as well as the Waymo Open Dataset. Code will be made publicly available.

We propose Point2Cyl, a supervised network transforming a raw 3D point cloud to a set of extrusion cylinders. Reverse engineering from a raw geometry to a CAD model is an essential task to enable manipulation of the 3D data in shape editing software and thus expand their usages in many downstream applications. Particularly, the form of CAD models having a sequence of extrusion cylinders --- a 2D sketch plus an extrusion axis and range --- and their boolean combinations is not only widely used in the CAD community/software but also has great expressivity of shapes, compared to having limited types of primitives (e.g., planes, spheres, and cylinders). In this work, we introduce a neural network that solves the extrusion cylinder decomposition problem in a geometry-grounded way by first learning underlying geometric proxies. Precisely, our approach first predicts per-point segmentation, base/barrel labels and normals, then estimates for the underlying extrusion parameters in differentiable and closed-form formulations. Our experiments show that our approach demonstrates the best performance on two recent CAD datasets, Fusion Gallery and DeepCAD, and we further showcase our approach on reverse engineering and editing.

This article addresses the problem of distilling knowledge from a large teacher model to a slim student network for LiDAR semantic segmentation. Directly employing previous distillation approaches yields inferior results due to the intrinsic challenges of point cloud, i.e., sparsity, randomness and varying density. To tackle the aforementioned problems, we propose the Point-to-Voxel Knowledge Distillation (PVD), which transfers the hidden knowledge from both point level and voxel level. Specifically, we first leverage both the pointwise and voxelwise output distillation to complement the sparse supervision signals. Then, to better exploit the structural information, we divide the whole point cloud into several supervoxels and design a difficultyaware sampling strategy to more frequently sample supervoxels containing less frequent classes and faraway objects. On these supervoxels, we propose inter-point and intervoxel affinity distillation, where the similarity information between points and voxels can help the student model better capture the structural information of the surrounding environment. We conduct extensive experiments on two popular LiDAR segmentation benchmarks, i.e., nuScenes [3] and SemanticKITTI [1]. On both benchmarks, our PVD consistently outperforms previous distillation approaches by a large margin on three representative backbones, i.e., Cylinder3D [27, 28], SPVNAS [20] and MinkowskiNet [5]. Notably, on the challenging nuScenes and SemanticKITTI datasets, our method can achieve roughly 75% MACs reduction and 2x speedup on the competitive Cylinder3D model and rank 1st on the SemanticKITTI leaderboard among all published algorithms. Our code is available at https://github.com/cardwing/Codes-for-PVKD.

Volumetric neural rendering methods like NeRF generate high-quality view synthesis results but are optimized per-scene leading to prohibitive reconstruction time. On the other hand, deep multi-view stereo methods can quickly reconstruct scene geometry via direct network inference. Point-NeRF combines the advantages of these two approaches by using neural 3D point clouds, with associated neural features, to model a radiance field. Point-NeRF can be rendered efficiently by aggregating neural point features near scene surfaces, in a ray marching-based rendering pipeline. Moreover, Point-NeRF can be initialized via direct inference of a pre-trained deep network to produce a neural point cloud; this point cloud can be fine-tuned to surpass the visual quality of NeRF with 30X faster training time. Point-NeRF can be combined with other 3D reconstruction methods and handles the errors and outliers in such methods via a novel pruning and growing mechanism.

In this work we present point-level region contrast, a self-supervised pre-training approach for the task of object detection. This approach is motivated by the two key factors in detection: localization and recognition. While accurate localization favors models that operate at the pixel- or point-level, correct recognition typically relies on a more holistic, region-level view of objects. Incorporating this perspective in pre-training, our approach performs contrastive learning by directly sampling individual point pairs from different regions. Compared to an aggregated representation per region, our approach is more robust to the change in input region quality, and further enables us to implicitly improve initial region assignments via online knowledge distillation during training. Both advantages are important when dealing with imperfect regions encountered in the unsupervised setting. Experiments show point-level region contrast improves on state-of-the-art pre-training methods for object detection and segmentation across multiple tasks and datasets, and we provide extensive ablation studies and visualizations to aid understanding. Code will be made available.

We present Point-BERT, a novel paradigm for learning Transformers to generalize the concept of BERT onto 3D point cloud. Following BERT, we devise a Masked Point Modeling (MPM) task to pre-train point cloud Transformers. Specifically, we first divide a point cloud into several local patches, and a point cloud Tokenizer is devised via a discrete Variational AutoEncoder (dVAE) to generate discrete point tokens containing meaningful local information. Then, we randomly mask some patches of input point clouds and feed them into the backbone Transformer. The pre-training objective is to recover the original point tokens at the masked locations under the supervision of point tokens obtained by the Tokenizer. Extensive experiments demonstrate that the proposed BERT-style pre-training strategy significantly improves the performance of standard point cloud Transformers. Equipped with our pre-training strategy, we show that a pure Transformer architecture attains 93.7% accuracy on ModelNet40 and 83.1% accuracy on the hardest setting of ScanObjectNN, surpassing carefully designed point cloud models with much fewer hand-made designs and human priors. We also demonstrate that the representations learned by Point-BERT transfer well to new tasks and domains, where our models largely advance the state-of-the-art of few-shot point cloud classification task.

LiDAR has become one of the primary 3D object detection sensors in autonomous driving. However, LiDAR's diverging point pattern with increasing distance results in a non-uniform sampled point cloud ill-suited to discretized volumetric feature extraction. Current methods either rely on voxelized point clouds or use inefficient farthest point sampling to mitigate detrimental effects caused by density variation but largely ignore point density as a feature and its predictable relationship with distance from the LiDAR sensor. Our proposed solution, Point Density-Aware Voxel network (PDV), is an end-to-end two stage LiDAR 3D object detection architecture that is designed to account for these point density variations. PDV efficiently localizes voxel features from the 3D sparse convolution backbone through voxel point centroids. The spatially localized voxel features are then aggregated through a density-aware RoI grid pooling module using kernel density estimation (KDE) and self-attention with point density positional encoding. Finally, we exploit LiDAR's point density to distance relationship to refine our final bounding box confidences. PDV outperforms all state-of-the-art methods on the Waymo Open Dataset and achieves competitive results on the KITTI dataset.

The construction of 3D point cloud datasets requires a great deal of human effort. Therefore, constructing a largescale 3D point clouds dataset is difficult. In order to remedy this issue, we propose a newly developed point cloud fractal database (PC-FractalDB), which is a novel family of formula-driven supervised learning inspired by fractal geometry encountered in natural 3D structures. Our research is based on the hypothesis that we could learn representations from more real-world 3D patterns than conventional 3D datasets by learning fractal geometry. We show how the PC-FractalDB facilitates solving several recent dataset-related problems in 3D scene understanding, such as 3D model collection and labor-intensive annotation. The experimental section shows how we achieved the performance rate of up to 61.9% and 59.0% for the ScanNetV2 and SUN RGB-D datasets, respectively, over the current highest scores obtained with the PointContrast, contrastive scene contexts (CSC), and RandomRooms. Moreover, the PC-FractalDB pre-trained model is especially effective in training with limited data. For example, in 10% of training data on ScanNetV2, the PC-FractalDB pre-trained VoteNet performs at 38.3%, which is +14.8% higher accuracy than CSC. Of particular note, we found that the proposed method achieves the highest results for 3D object detection pre-training in limited point cloud data.

In this paper, we present Point Cloud Color Constancy, in short PCCC, an illumination chromaticity estimation algorithm exploiting a point cloud. We leverage the depth information captured by the time-of-flight (ToF) sensor mounted rigidly with the RGB sensor, and form a 6D cloud where each point contains the coordinates and RGB intensities, noted as (x,y,z,r,g,b). PCCC applies the PointNet architecture to the color constancy problem, deriving the illumination vector point-wise and then making a global decision about the global illumination chromaticity. On two popular RGB-D datasets, which we extend with illumination information, as well as on a novel benchmark, PCCC obtains lower error than the state-of-the-art algorithms. Our method is simple and fast, requiring merely 16*16-size input and reaching speed over 500 fps, including the cost of building the point cloud and net inference.