CVPR 2021 Papers

We present SOLID-Net, a neural network for spatially-varying outdoor lighting estimation from a single outdoor image for any 2D pixel location. Previous work has used a unified sky environment map to represent outdoor lighting. Instead, we generate spatially-varying local lighting environment maps by combining global sky environment map with warped image information according to geometric information estimated from intrinsics. As no outdoor dataset with image and local lighting ground truth is readily available, we introduce SOLID-Img dataset with physically-based rendered images and their corresponding intrinsic and lighting information. We train a deep neural network to regress intrinsic cues with physically-based constrains and use them to conduct global and local lightings estimation. Experiments on both synthetic and real datasets show that SOLID-Net significantly outperforms previous methods.

Transferring makeup from the misaligned reference image is challenging. Previous methods overcome this barrier by computing pixel-wise correspondences between two images, which is inaccurate and computational-expensive. In this paper, we take a different perspective to break down the makeup transfer problem into a two-step extraction-assignment process. To this end, we propose a Style-based Controllable GAN model that consists of three components, each of which corresponds to target style-code encoding, face identity features extraction, and makeup fusion, respectively. In particular, a Part-specific Style Encoder encodes the component-wise makeup style of the reference image into a style-code in an intermediate latent space W. The style-code discards spatial information and therefore is invariant to spatial misalignment. On the other hand, the style-code embeds component-wise information, enabling flexible partial makeup editing from multiple references. This style-code, together with source identity features, are integrated to a Makeup Fusion Decoder equipped with multiple AdaIN layers to generate the final result. Our proposed method demonstrates great flexibility on makeup transfer by supporting makeup removal, shade-controllable makeup transfer, and part-specific makeup transfer, even with large spatial misalignment. Extensive experiments demonstrate the superiority of our approach over state-of-the-art methods. Code is available at https://github.com/makeuptransfer/SCGAN.

We introduce a new generator architecture, aimed at fast and efficient high-resolution image-to-image translation. We design the generator to be an extremely lightweight function of the full-resolution image. In fact, we use pixel-wise networks; that is, each pixel is processed independently of others, through a composition of simple affine transformations and nonlinearities. We take three important steps to equip such a seemingly simple function with adequate expressivity. First, the parameters of the pixel-wise networks are spatially varying so they can represent a broader function class than simple 1x1 convolutions. Second, these parameters are predicted by a fast convolutional network that processes an aggressively low-resolution representation of the input. Third, we augment the input image with a sinusoidal encoding of spatial coordinates, which provides an effective inductive bias for generating realistic novel high-frequency image content. As a result, our model is up to 18x faster than state-of-the-art baselines. We achieve this speedup while generating comparable visual quality across different image resolutions and translation domains.

Self-supervised learning has been widely used to obtain transferrable representations from unlabeled images. Especially, recent contrastive learning methods have shown impressive performances on downstream image classification tasks. While these contrastive methods mainly focus on generating invariant global representations at the image-level under semantic-preserving transformations, they are prone to overlook spatial consistency of local representations and therefore have a limitation in pretraining for localization tasks such as object detection and instance segmentation. Moreover, aggressively cropped views used in existing contrastive methods can minimize representation distances between the semantically different regions of a single image. In this paper, we propose a spatially consistent representation learning algorithm (SCRL) for multi-object and location-specific tasks. In particular, we devise a novel self-supervised objective that tries to produce coherent spatial representations of a randomly cropped local region according to geometric translations and zooming operations. On various downstream localization tasks with benchmark datasets, the proposed SCRL shows significant performance improvements over the image-level supervised pretraining as well as the state-of-the-art self-supervised learning methods.

Video-based person re-identification aims to match pedestrians from video sequences across non-overlapping camera views. The key factor for video person re-identification is to effectively exploit both spatial and temporal clues from video sequences. In this work, we propose a novel Spatial-Temporal Correlation and Topology Learning framework (CTL) to pursue discriminative and robust representation by modeling cross-scale spatial-temporal correlation. Specifically, CTL utilizes a CNN backbone and a key-points estimator to extract semantic local features from human body at multiple granularities as graph nodes. It explores a context-reinforced topology to construct multi-scale graphs by considering both global contextual information and physical connections of human body. Moreover, a 3D graph convolution and a cross-scale graph convolution are designed, which facilitate direct cross-spacetime and cross-scale information propagation for capturing hierarchical spatial-temporal dependencies and structural information. By jointly performing the two convolutions, CTL effectively mines comprehensive clues that are complementary with appearance information to enhance representational capacity. Extensive experiments on two video benchmarks have demonstrated the effectiveness of the proposed method and the state-of-the-art performance.

The remarkable success in face forgery techniques has received considerable attention in computer vision due to security concerns. We observe that up-sampling is a necessary step of most face forgery techniques, and cumulative up-sampling will result in obvious changes in the frequency domain, especially in the phase spectrum. According to the property of natural images, the phase spectrum preserves abundant frequency components that provide extra information and complement the loss of the amplitude spectrum. To this end, we present a novel Spatial-Phase Shallow Learning (SPSL) method, which combines spatial image and phase spectrum to capture the up-sampling artifacts of face forgery to improve the transferability, for face forgery detection. And we also theoretically analyze the validity of utilizing the phase spectrum. Moreover, we notice that local texture information is more crucial than high-level semantic information for the face forgery detection task. So we reduce the receptive fields by shallowing the network to suppress high-level features and focus on the local region. Extensive experiments show that SPSL can achieve the state-of-the-art performance on cross-datasets evaluation as well as multi-class classification and obtain comparable results on single dataset evaluation.

Modern one-stage video instance segmentation networks suffer from two limitations. First, convolutional features are neither aligned with anchor boxes nor with ground-truth bounding boxes, reducing the mask sensitivity to spatial location. Second, a video is directly divided into individual frames for frame-level instance segmentation, ignoring the temporal correlation between adjacent frames. To address these issues, we propose a simple yet effective one-stage video instance segmentation framework by spatial calibration and temporal fusion, namely STMask. To ensure spatial feature calibration with ground-truth bounding boxes, we first predict regressed bounding boxes around ground-truth bounding boxes, and extract features from them for frame-level instance segmentation. To further explore temporal correlation among video frames, we aggregate a temporal fusion module to infer instance masks from each frame to its adjacent frames, which helps our framework to handle challenging videos such as motion blur, partial occlusion and unusual object-to-camera poses. Experiments on the YouTube-VIS valid set show that the proposed STMask with ResNet-50/-101 backbone obtains 33.5 % / 36.8 % mask AP, while achieving 28.6 / 23.4 FPS on video instance segmentation. The code is released online https://github.com/MinghanLi/STMask.

It has been long recognized that deep neural networks are sensitive to changes in spatial configurations or scene structures. Image augmentations, such as random translation, cropping, and resizing, can be used to improve the robustness of deep neural networks under spatial transforms. However, changes in object part configurations, spatial layout of object, and scene structures of the images may still result in major changes in the their feature representations generated by the network, creating significant challenges for various visual learning tasks, including representation or metric learning, image classification and retrieval. In this work, we introduce a new learnable module, called spatial assembly network (SAN), to address this important issue. This SAN module examines the input image and performs a learned re-organization and assembly of feature points from different spatial locations conditioned by feature maps from previous network layers so as to maximize the discriminative power of the final feature representation. This differentiable module can be flexibly incorporated into existing network architectures, improving their capabilities in handling spatial variations and structural changes of the image scene. We demonstrate that the proposed SAN module is able to significantly improve the performance of various metric / representation learning, image retrieval and classification tasks, in both supervised and unsupervised learning scenarios.

We present Sparse R-CNN, a purely sparse method for object detection in images. Existing works on object detection heavily rely on dense object candidates, such as k anchor boxes pre-defined on all grids of image feature map of size HxW. In our method, however, a fixed sparse set of learned object proposals, total length of N, are provided to object recognition head to perform classification and location. By eliminating HWk (up to hundreds of thousands) hand-designed object candidates to N (e.g. 100) learnable proposals, Sparse R-CNN completely avoids all efforts related to object candidates design and many-to-one label assignment. More importantly, final predictions are directly output without non-maximum suppression post-procedure. Sparse R-CNN demonstrates accuracy, run-time and training convergence performance on par with the well-established detector baselines on the challenging COCO dataset, e.g., achieving 45.0 AP in standard 3x training schedule and running at 22 fps using ResNet-50 FPN model. We hope our work could inspire re-thinking the convention of dense prior in object detectors. The code is available at: https://github.com/PeizeSun/SparseR-CNN.

Three-dimensional scanning by means of structured light illumination is an active imaging technique involving projecting and capturing a series of striped patterns and then using the observed warping of stripes to reconstruct the target object's surface through triangulating each pixel in the camera to a unique projector coordinate corresponding to a particular feature in the projected patterns. The undesirable phenomenon of multi-path occurs when a camera pixel simultaneously sees features from multiple projector coordinates. Bimodal multi-path is a particularly common situation found along step edges, where the camera pixel sees both a foreground and background surface. Generalized from bimodal multi-path, this paper looks at sparse or N modal multi-path as a more general case, where the camera pixel sees no less than two reflective surfaces, resulting in decoding errors. Using fringe projection profilometry, our proposed solution is to treat each camera pixel as an underdetermined linear system of equations and to find the sparsest (least number of paths) solution using an application-specific Bayesian learning approach. We validate this algorithm with both simulations and a number of challenging real-world scenarios, outperforming the state-of-the-art techniques.

Estimating scene geometry from cost-effective sensors is key for robots. In this paper, we study the problem of predicting dense depth from a single RGB image (monodepth) with optional sparse measurements from low-cost active depth sensors. We introduce Sparse Auxiliary Networks (SAN), a new module enabling monodepth networks to perform both the tasks of depth prediction and completion, depending on whether only RGB images or also sparse point clouds are available at inference time. First, we decouple the image and depth map encoding stages using sparse convolutions to process only the valid depth map pixels. Second, we inject this information, when available, into the skip connections of the depth prediction network, augmenting its features. Through extensive experimental analysis on one indoor (NYUv2) and two outdoor (KITTI and DDAD) benchmarks, we demonstrate that our proposed SAN architecture is able to simultaneously learn both tasks, while achieving a new state of the art in depth prediction by a significant margin.

We present a method that learns a spatiotemporal neural irradiance field for dynamic scenes from a single video. Our learned representation enables free-viewpoint rendering of the input video. Our method builds upon recent advances in implicit representations. Learning a spatiotemporal irradiance field from a single video poses significant challenges because the video contains only one observation of the scene at any point in time. The 3D geometry of a scene can be legitimately represented in numerous ways since varying geometry (motion) can be explained with varying appearance and vice versa. We address this ambiguity by constraining the time-varying geometry of our dynamic scene representation using the scene depth estimated from video depth estimation methods, aggregating contents from individual frames into a single global representation. We provide an extensive quantitative evaluation and demonstrate compelling free-viewpoint rendering results.

Compact video super-resolution (VSR) networks can be easily deployed on resource-limited devices, e.g., smart-phones and wearable devices, but have considerable performance gaps compared with complicated VSR networks that require a large amount of computing resources. In this paper, we aim to improve the performance of compact VSR networks without changing their original architectures, through a knowledge distillation approach that transfers knowledge from a complicated VSR network to a compact one. Specifically, we propose a space-time distillation (STD) scheme to exploit both spatial and temporal knowledge in the VSR task. For space distillation, we extract spatial attention maps that hints the high-frequency video content from both networks, which are further used for transferring spatial modeling ability. For time distillation, we narrow the performance gap between compact models and complicated models by distilling the feature similarity of the temporal memory cells, which is encoded from the sequence of feature maps generated in the training clips using ConvLSTM. During the training process, STD can be easily incorporated into any network without changing the original network architecture. Experimental results on standard benchmarks demonstrate that, in resource-constrained situations, the proposed method notably improve the performance of existing VSR networks without increasing the inference time.

Unsupervised Domain Adaptation (UDA) can tackle the challenge that convolutional neural network (CNN)-based approaches for semantic segmentation heavily rely on the pixel-level annotated data, which is labor-intensive. However, existing UDA approaches in this regard inevitably require the full access to source datasets to reduce the gap between the source and target domains during model adaptation, which are impractical in the real scenarios where the source datasets are private, and thus cannot be released along with the well-trained source models. To cope with this issue, we propose a source-free domain adaptation framework for semantic segmentation, namely SFDA, in which only a well-trained source model and an unlabeled target domain dataset are available for adaptation. SFDA not only enables to recover and preserve the source domain knowledge from the source model via knowledge transfer during model adaptation, but also distills valuable information from the target domain for self-supervised learning. The pixel- and patch-level optimization objectives tailored for semantic segmentation are seamlessly integrated in the framework. The extensive experimental results on numerous benchmark datasets highlight the effectiveness of our framework against the existing UDA approaches relying on source data.

Federated learning (FL) is a popular distributed learning framework that can reduce privacy risks by not explicitly sharing private data. However, recent works have demonstrated that sharing model updates makes FL vulnerable to inference attack. In this work, we show our key observation that the data representation leakage from gradients is the essential cause of privacy leakage in FL. We also provide an analysis of this observation to explain how the data presentation is leaked. Based on this observation, we propose a defense called Soteria against model inversion attack in FL. The key idea of our defense is learning to perturb data representation such that the quality of the reconstructed data is severely degraded, while FL performance is maintained. In addition, we derive a certified robustness guarantee to FL and a convergence guarantee to FedAvg, after applying our defense. To evaluate our defense, we conduct experiments on MNIST and CIFAR10 for defending against the DLG attack and GS attack. Without sacrificing accuracy, the results demonstrate that our proposed defense can increase the mean squared error between the reconstructed data and the raw data by as much as 160x for both DLG attack and GS attack, compared with baseline defense methods. Therefore, the privacy of the FL system is significantly improved. Our code can be found at https://github.com/jeremy313/Soteria.

The recently introduced introspective variational autoencoder (IntroVAE) exhibits outstanding image generations, and allows for amortized inference using an image encoder. The main idea in IntroVAE is to train a VAE adversarially, using the VAE encoder to discriminate between generated and real data samples. However, the original IntroVAE loss function relied on a particular hinge-loss formulation that is very hard to stabilize in practice, and its theoretical convergence analysis ignored important terms in the loss. In this work, we take a step towards better understanding of the IntroVAE model, its practical implementation, and its applications. We propose the Soft-IntroVAE, a modified IntroVAE that replaces the hinge-loss terms with a smooth exponential loss on generated samples. This change significantly improves training stability, and also enables theoretical analysis of the complete algorithm. Interestingly, we show that the IntroVAE converges to a distribution that minimizes a sum of KL distance from the data distribution and an entropy term. We discuss the implications of this result, and demonstrate that it induces competitive image generation and reconstruction. Finally, we describe an application of Soft-IntroVAE to unsupervised image translation, and demonstrate compelling results. Code and additional information is available on the project website - taldatech.github.io/soft-intro-vae-web

Image-to-Image (I2I) multi-domain translation models are usually evaluated also using the quality of their semantic interpolation results. However, state-of-the-art models frequently show abrupt changes in the image appearance during interpolation, and usually perform poorly in interpolations across domains. In this paper, we propose a new training protocol based on three specific losses which help a translation network to learn a smooth and disentangled latent style space in which: 1) Both intra- and inter-domain interpolations correspond to gradual changes in the generated images and 2) The content of the source image is better preserved during the translation. Moreover, we propose a novel evaluation metric to properly measure the smoothness of latent style space of I2I translation models. The proposed method can be plugged in existing translation approaches, and our extensive experiments on different datasets show that it can significantly boost the quality of the generated images and the graduality of the interpolations.

Neural image compression leverages deep neural networks to outperform traditional image codecs in rate-distortion performance. However, the resulting models are also heavy, computationally demanding and generally optimized for a single rate, limiting their practical use. Focusing on practical image compression, we propose slimmable compressive autoencoders (SlimCAEs), where rate (R) and distortion (D) are jointly optimized for different capacities. Once trained, encoders and decoders can be executed at different capacities, leading to different rates and complexities. We show that a successful implementation of SlimCAEs requires suitable capacity-specific RD tradeoffs. Our experiments show that SlimCAEs are highly flexible models that provide excellent rate-distortion performance, variable rate, and dynamic adjustment of memory, computational cost and latency, thus addressing the main requirements of practical image compression.

We introduce a novel deep neural network to estimate a depth map from a single monocular indoor panorama. The network directly works on the equirectangular projection, exploiting the properties of indoor 360 images. Starting from the fact that gravity plays an important role in the design and construction of man-made indoor scenes, we propose a compact representation of the scene into vertical slices of the sphere, and we exploit long- and short-term relationships among slices to recover the equirectangular depth map. Our design makes it possible to maintain high-resolution information in the extracted features even with a deep network. The experimental results demonstrate that our method outperforms current state-of-the-art solutions in prediction accuracy, particularly for real-world data.