CVPR 2022 Papers

Lidars and cameras are critical sensors that provide complementary information for 3D detection in autonomous driving. While prevalent multi-modal methods simply decorate raw lidar point clouds with camera features and feed them directly to existing 3D detection models, our study shows that fusing camera features with deep lidar features instead of raw points, can lead to better performance. However, as those features are often augmented and aggregated, a key challenge in fusion is how to effectively align the transformed features from two modalities. In this paper, we propose two novel techniques: InverseAug that inverses geometric-related augmentations, e.g., rotation, to enable accurate geometric alignment between lidar points and image pixels, and LearnableAlign that leverages cross-attention to dynamically capture the correlations between image and lidar features during fusion. Based on InverseAug and LearnableAlign, we develop a family of generic multi-modal 3D detection models named DeepFusion, which is more accurate than previous methods. For example, DeepFusion improves PointPillars, CenterPoint, and 3D-MAN baselines on Pedestrian detection for 6.7, 8.9, and 6.2 LEVEL_2 APH, respectively. Notably, our models achieve state-of-the-art performance on Waymo Open Dataset, and show strong model robustness against input corruptions and out-of-distribution data. Code will be publicly available at https://github.com/tensorflow/lingvo.

In recent years, with the advances of generative models, many powerful face manipulation systems have been developed based on Deep Neural Networks (DNNs), called DeepFakes. If DeepFakes are not controlled timely and properly, they would become a real threat to both celebrities and ordinary people. Precautions such as adding perturbations to the source inputs will make DeepFake results look distorted from the perspective of human eyes. However, previous method doesn't explore whether the disrupted images can still spoof DeepFake detectors. This is critical for many applications where DeepFake detectors are used to discriminate between DeepFake data and real data due to the huge cost of examining a large amount of data manually. We argue that the detectors do not share a similar perspective as human eyes, which might still be spoofed by the disrupted data. Besides, the existing disruption methods rely on iteration-based perturbation generation algorithms, which is time-consuming. In this paper, we propose a novel DeepFake disruption algorithm called "DeepFake Disrupter". By training a perturbation generator, we can add the human-imperceptible perturbations to source images that need to be protected without any backpropagation update. The DeepFake results of these protected source inputs would not only look unrealistic by the human eye but also can be distinguished by DeepFake detectors easily. For example, experimental results show that by adding our trained perturbations, fake images generated by StarGAN can result in a 10 20% increase in F1-score evaluated by various DeepFake detectors.

Face identification (FI) is ubiquitous and drives many high-stake decisions made by the law enforcement. State-of-the-art FI approaches compare two images by taking the cosine similarity between their image embeddings. Yet, such approach suffers from poor out-of-distribution (OOD) generalization to new types of images (e.g., when a query face is masked, cropped or rotated) not included in the training set or the gallery. Here, we propose a re-ranking approach that compares two faces using the Earth Mover's Distance on the deep, spatial features of image patches. Our extra comparison stage explicitly examines image similarity at a fine-grained level (e.g., eyes to eyes) and is more robust to OOD perturbations and occlusions than traditional FI. Interestingly, without finetuning feature extractors, our method consistently improves the accuracy on all tested OOD queries: masked, cropped, rotated, and adversarial while obtaining similar results on in-distribution images.

Deep Learning (DL) has shown great promise in the unsupervised task of clustering. That said, while in classical (i.e., non-deep) clustering the benefits of the nonparametric approach are well known, most deep-clustering methods are parametric: namely, they require a predefined and fixed number of clusters, denoted by K. When K is unknown, however, using model-selection criteria to choose its optimal value might become computationally expensive, especially in DL as the training process would have to be repeated numerous times. In this work, we bridge this gap by introducing an effective deep-clustering method that does not require knowing the value of K as it infers it during the learning. Using a split/merge framework, a dynamic architecture that adapts to the changing K, and a novel loss, our proposed method outperforms existing nonparametric methods (both classical and deep ones). While the very few existing deep nonparametric methods lack scalability, we demonstrate ours by being the first to report the performance of such a method on ImageNet. We also demonstrate the importance of inferring K by showing how methods that fix it deteriorate in performance when their assumed K value gets further from the ground-truth one, especially on imbalanced datasets. Our code is available at https://github.com/BGU-CS-VIL/DeepDPM.

Recent techniques have been successful in reconstructing surfaces as level sets of learned functions (such as signed distance fields) parameterized by deep neural networks. Many of these methods, however, learn only closed surfaces and are unable to reconstruct shapes with boundary curves. We propose a hybrid shape representation that combines explicit boundary curves with implicit learned interiors. Using machinery from geometric measure theory, we parameterize currents using deep networks and use stochastic gradient descent to solve a minimal surface problem. By modifying the metric according to target geometry coming, e.g., from a mesh or point cloud, we can use this approach to represent arbitrary surfaces, learning implicitly defined shapes with explicitly defined boundary curves. We further demonstrate learning families of shapes jointly parameterized by boundary curves and latent codes.

Deep learning has improved vanishing point detection in images. Yet, deep networks require expensive annotated datasets trained on costly hardware and do not generalize to even slightly different domains, and minor problem variants. Here, we address these issues by injecting deep vanishing point detection networks with prior knowledge. This prior knowledge no longer needs to be learned from data, saving valuable annotation efforts and compute, unlocking realistic few-sample scenarios, and reducing the impact of domain changes. Moreover, the interpretability of the priors allows to adapt deep networks to minor problem variations such as switching between Manhattan and non-Manhattan worlds. We seamlessly incorporate two geometric priors: (i) Hough Transform -- mapping image pixels to straight lines, and (ii) Gaussian sphere -- mapping lines to great circles whose intersections denote vanishing points. Experimentally, we ablate our choices and show comparable accuracy to existing models in the large-data setting. We validate our model's improved data efficiency, robustness to domain changes, adaptability to non-Manhattan settings.

Recent legislation has led to interest in machine unlearning, i.e., removing specific training samples from a predictive model as if they never existed in the training dataset. Unlearning may also be required due to corrupted/adversarial data or simply a user's updated privacy requirement. For models which require no training (k-NN), simply deleting the closest original sample can be effective. But this idea is inapplicable to models which learn richer representations. Recent ideas leveraging optimization-based updates scale poorly with the model dimension d, due to inverting the Hessian of the loss function. We use a variant of a new conditional independence coefficient, L-CODEC, to identify a subset of the model parameters with the most semantic overlap on an individual sample level. Our approach completely avoids the need to invert a (possibly) huge matrix. By utilizing a Markov blanket selection, we premise that L-CODEC is also suitable for deep unlearning, as well as other applications in vision. Compared to alternatives, L-CODEC makes approximate unlearning possible in settings that would otherwise be infeasible, including vision models used for face recognition, person re-identification and NLP models that may require unlearning samples identified for exclusion. Code can be found at https://github.com/vsingh-group/LCODEC-deep-unlearning/

Existing learning-based stereo compression methods usually adopt a unidirectional approach to encoding one image independently and the other image conditioned upon the first. This paper proposes a novel bi-directional coding-based end-to-end stereo image compression network (BCSIC-Net). BCSIC-Net consists of a novel bi-directional contextual transform module which performs nonlinear transform conditioned upon the inter-view context in a latent space to reduce inter-view redundancy, and a bi-directional conditional entropy model that employs inter-view correspondence as a conditional prior to improve coding efficiency. Experimental results on the InStereo2K and KITTI datasets demonstrate that the proposed BCSIC-Net can effectively reduce the inter-view redundancy and outperforms state-of-the-art methods.

Using only a model that was trained to predict where people look at images, and no additional training data, we can produce a range of powerful editing effects for reducing distraction in images. Given an image and a mask specifying the region to edit, we backpropagate through a state-of-the-art saliency model to parameterize a differentiable editing operator, such that the saliency within the masked region is reduced. We demonstrate several operators, including: a recoloring operator, which learns to apply a color transform that camouflages and blends distractors into their surroundings; a warping operator, which warps less salient image regions to cover distractors, gradually collapsing objects into themselves and effectively removing them (an effect akin to inpainting); a GAN operator, which uses a semantic prior to fully replace image regions with plausible, less salient alternatives. The resulting effects are consistent with cognitive research on the human visual system (e.g., since color mismatch is salient, the recoloring operator learns to harmonize objects' colors with their surrounding to reduce their saliency), and, importantly, are all achieved solely through the guidance of the pretrained saliency model. We present results on a variety of natural images and conduct a perceptual study to evaluate and validate the changes in viewers' eye-gaze between the original images and our edited results.

Multi-view clustering has been shown to boost clustering performance by effectively mining the complementary information from multiple views. However, we observe that learning from data with more views is not guaranteed to achieve better clustering performance than from data with fewer views. To address this issue, we propose a general deep learning based framework that is guaranteed to reduce the risk of performance degradation caused by view increase. Concretely, the model is trained to simultaneously extract complementary information and discard the meaningless noise by automatically selecting features. These two learning procedures are incorporated into one unified framework by the proposed optimization objective. In theory, the empirical clustering risk of the model is no higher than learning from data before the view increase and data of the new increased single view. Also, the expected clustering risk of the model under divergence-based loss is no higher than that with high probability. Comprehensive experiments on benchmark datasets demonstrate the effectiveness and superiority of the proposed framework in achieving safe multi-view clustering.

State-of-the-art fully intrinsic network for non-rigid shape matching are unable to disambiguate between shape inner symmetries. Meanwhile, recent advances in the functional map framework allow to enforce orientation preservation using a functional representation for tangent vector field transfer, through so-called complex functional maps. Using this representation, we propose a new deep learning approach to learn orientation-aware features in a fully unsupervised setting. Our architecture is built on DiffusionNet, which makes our method robust to discretization changes, while adding a vector-field-based loss, which promotes orientation preservation without using (often unstable) extrinsic descriptors. Our source code is available at: https://github.com/nicolasdonati/DUO-FM

Integrating a foreground object into a background scenewith illumination harmonization is an important but chal-lenging task in computer vision and augmented reality community. Existing methods mainly focus on foreground andbackground appearance consistency or the foreground object shadow generation, which rarely consider global appearance and illumination harmonization. In this paper,we formulate seamless illumination harmonization as anillumination exchange and aggregation problem. Specifi-cally, we firstly apply a physically-based rendering methodto construct a large-scale, high-quality dataset (named IH)for our task, which contains various types of foreground ob-jects and background scenes with different lighting conditions. Then, we propose a deep image-based illuminationharmonization GAN framework named DIH-GAN, whichmakes full use of a multi-scale attention mechanism and illumination exchange strategy to directly infer mapping rela-tionship between the inserted foreground object and the corresponding background scene. Meanwhile, we also use adversarial learning strategy to further refine the illuminationharmonization result. Our method can not only achieve har-monious appearance and illumination for the foregroundobject but also can generate compelling shadow cast bythe foreground object. Comprehensive experiments on bothour IH dataset and real-world images show that our pro-posed DIH-GAN provides a practical and effective solutionfor image-based object illumination harmonization editing,and validate the superiority of our method against state-of-the-art methods.

Depth reconstruction and hyperspectral reflectance reconstruction are two active research topics in computer vision and image processing. Conventionally, these two topics have been studied separately using independent imaging setups and there is no existing method which can acquire depth and spectral reflectance simultaneously in one shot without using special hardware. In this paper, we propose a novel single-shot hyperspectral-depth reconstruction method using an off-the-shelf RGB camera and projector. Our method is based on a single color-dot projection, which simultaneously acts as structured light for depth reconstruction and spatially-varying color illuminations for hyperspectral reflectance reconstruction. To jointly reconstruct the depth and the hyperspectral reflectance from a single color-dot image, we propose a novel end-to-end network architecture that effectively incorporates a geometric color-dot pattern loss and a photometric hyperspectral reflectance loss. Through the experiments, we demonstrate that our hyperspectral-depth reconstruction method outperforms the combination of an existing state-of-the-art single-shot hyperspectral reflectance reconstruction method and depth reconstruction method.

We propose a principled and practical method for out-of-distribution (OoD) detection with deep hybrid models (DHMs), which model the joint density p(x,y) of features and labels with a single forward pass. By factorizing the joint density p(x,y) into three sources of uncertainty, we show that our approach has the ability to identify samples semantically different from the training data. To ensure computational scalability, we add a weight normalization step during training, which enables us to plug in state-of-the-art (SoTA) deep neural network (DNN) architectures for approximately modeling and inferring expressive probability distributions. Our method provides an efficient, general, and flexible framework for predictive uncertainty estimation with promising results and theoretical support. To our knowledge, this is the first work to reach 100% in OoD detection tasks on both vision and language datasets, especially on notably difficult dataset pairs such as CIFAR-10 vs. SVHN and CIFAR-100 vs. CIFAR-10. This work is a step towards enabling DNNs in real-world deployment for safety-critical applications.

Humans are able to recognize structured relations in observation, allowing us to decompose complex scenes into simpler parts and abstract the visual world in multiple levels. However, such hierarchical reasoning ability of human perception remains largely unexplored in current literature of semantic segmentation. Existing work is often aware of flatten labels and predicts target classes exclusively for each pixel. In this paper, we instead address hierarchical semantic segmentation (HSS), which aims at structured, pixel-wise description of visual observation in terms of a class hierarchy. We devise HSSN, a general HSS framework that tackles two critical issues in this task: i) how to efficiently adapt existing hierarchy-agnostic segmentation networks to the HSS setting, and ii) how to leverage the hierarchy information to regularize HSS network learning. To address i), HSSN directly casts HSS as a pixel-wise multi-label classification task, only bringing minimal architecture change to current segmentation models. To solve ii), HSSN first explores inherent properties of the hierarchy as a training objective, which enforces segmentation predictions to obey the hierarchy structure. Further, with hierarchy-induced margin constraints, HSSN reshapes the pixel embedding space, so as to generate well-structured pixel representations and improve segmentation eventually. We conduct experiments on four semantic segmentation datasets (i.e., Mapillary Vistas 2.0, Cityscapes, LIP, and PASCAL-Person-Part), with different class hierarchies, segmentation network architectures and backbones, showing the generalization and superiority of HSSN.

Deep neural networks (DNN) have achieved great success in image restoration. However, most DNN methods are designed as a black box, lacking transparency and interpretability. Although some methods are proposed to combine traditional optimization algorithms with DNN, they usually demand pre-defined degradation processes or handcrafted assumptions, making it difficult to deal with complex and real-world applications. In this paper, we propose a Deep Generalized Unfolding Network (DGUNet) for image restoration. Concretely, without loss of interpretability, we integrate a gradient estimation strategy into the gradient descent step of the Proximal Gradient Descent (PGD) algorithm, driving it to deal with complex and real-world image degradation. In addition, we design inter-stage information pathways across proximal mapping in different PGD iterations to rectify the intrinsic information loss in most deep unfolding networks (DUN) through a multi-scale and spatial-adaptive way. By integrating the flexible gradient descent and informative proximal mapping, we unfold the iterative PGD algorithm into a trainable DNN. Extensive experiments on various image restoration tasks demonstrate the superiority of our method in terms of state-of-the-art performance, interpretability, and generalizability. The source code is available at https://github.com/MC-E/Deep-Generalized-Unfolding-Networks-for-Image-Restoration.

Many recent state-of-the-art (SOTA) optical flow models use finite-step recurrent update operations to emulate traditional algorithms by encouraging iterative refinements toward a stable flow estimation. However, these RNNs impose large computation and memory overheads, and are not directly trained to model such "stable estimation". They can converge poorly and thereby suffer from performance degradation. To combat these drawbacks, we propose deep equilibrium (DEQ) flow estimators, an approach that directly solves for the flow as the infinite-level fixed point of an implicit layer (using any black-box solver), and differentiates through this fixed point analytically (thus requiring O(1) training memory). This implicit-depth approach is not predicated on any specific model, and thus can be applied to a wide range of SOTA flow estimation model designs (e.g., RAFT and GMA). The use of these DEQ flow estimators allows us to compute the flow faster using, e.g., fixed-point reuse and inexact gradients, consumes 4-6x less training memory than the recurrent counterpart, and achieves better results with the same computation budget. In addition, we propose a novel, sparse fixed-point correction scheme to stabilize our DEQ flow estimators, which addresses a longstanding challenge for DEQ models in general. We test our approach in various realistic settings and show that it improves SOTA methods on Sintel and KITTI datasets with substantially better computational and memory efficiency.

Depth-from-focus (DFF) is a technique that infers depth using the focus change of a camera. In this work, we propose a convolutional neural network (CNN) to find the best-focused pixels in a focal stack and infer depth from the focus estimation. The key innovation of the network is the novel deep differential focus volume (DFV). By computing the first-order derivative with the stacked features over different focal distances, DFV is able to capture both the focus and context information for focus analysis. Besides, we also introduce a probability regression mechanism for focus estimation to handle sparsely sampled focal stacks and provide uncertainty estimation to the final prediction. Comprehensive experiments demonstrate that the proposed model achieves state-of-the-art performance on multiple datasets with good generalizability and fast speed.

In this paper, we tackle the problem of blind image super-resolution(SR) with a reformulated degradation model and two novel modules. Following the common practices of blind SR, our method proposes to improve both the kernel estimation as well as the kernel-based high-resolution image restoration. To be more specific, we first reformulate the degradation model such that the deblurring kernel estimation can be transferred into the low-resolution space. On top of this, we introduce a dynamic deep linear filter module. Instead of learning a fixed kernel for all images, it can adaptively generate deblurring kernel weights conditional on the input and yield a more robust kernel estimation. Subsequently, a deep constrained least square filtering module is applied to generate clean features based on the reformulation and estimated kernel. The deblurred feature and the low input image feature are then fed into a dual-path structured SR network and restore the final high-resolution result. To evaluate our method, we further conduct evaluations on several benchmarks, including Gaussian8 and DIV2KRK. Our experiments demonstrate that the proposed method achieves better accuracy and visual improvements against state-of-the-art methods. Codes and models are available at https://github.com/megvii-research/DCLS-SR.