CVPR 2022 Papers

Single image super-resolution (SISR) with generative adversarial networks (GAN) has recently attracted increasing attention due to its potentials to generate rich details. However, the training of GAN is unstable, and it often introduces many perceptually unpleasant artifacts along with the generated details. In this paper, we demonstrate that it is possible to train a GAN-based SISR model which can stably generate perceptually realistic details while inhibiting visual artifacts. Based on the observation that the local statistics (e.g., residual variance) of artifact areas are often different from the areas of perceptually friendly details, we develop a framework to discriminate between GAN-generated artifacts and realistic details, and consequently generate an artifact map to regularize and stabilize the model training process. Our proposed locally discriminative learning (LDL) method is simple yet effective, which can be easily plugged in off-the-shelf SISR methods and boost their performance. Experiments demonstrate that LDL outperforms the state-of-the-art GAN based SISR methods, achieving not only higher reconstruction accuracy but also superior perceptual quality on both synthetic and real-world datasets. Codes and models are available at https://github.com/csjliang/LDL.

We describe a method to extract persistent elements of a dynamic scene from an input video. We represent each scene element as a Deformable Sprite consisting of three components: 1) a 2D texture image for the entire video, 2) per-frame masks for the element, and 3) non-rigid deformations that map the texture image into each video frame. The resulting decomposition allows for applications such as consistent video editing. Deformable Sprites are a type of video auto-encoder model that is optimized on individual videos, and does not require training on a large dataset, nor does it rely on pre-trained models. Moreover, our method does not require object masks or other user input, and discovers moving objects of a wider variety than previous work. We evaluate our approach on standard video datasets and show qualitative results on a diverse array of Internet videos.

In this paper, we propose a new open-source benchmarking framework for Visual Geo-localization (VG) that allows to build, train, and test a wide range of commonly used architectures, with the flexibility to change individual components of a geo-localization pipeline. The purpose of this framework is twofold: i) gaining insights into how different components and design choices in a VG pipeline impact the final results, both in terms of performance (recall@N metric) and system requirements (such as execution time and memory consumption); ii) establish a systematic evaluation protocol for comparing different methods. Using the proposed framework, we perform a large suite of experiments which provide criteria for choosing backbone, aggregation and negative mining depending on the use-case and requirements. We also assess the impact of engineering techniques like pre/post-processing, data augmentation and image resizing, showing that better performance can be obtained through somewhat simple procedures: for example, downscaling the images' resolution to 80% can lead to similar results with a 36% savings in extraction time and dataset storage requirement. Code and trained models are available at https://deep-vg-bench.herokuapp.com/.

Unsupervised localization and segmentation are long-standing computer vision challenges that involve decomposing an image into semantically-meaningful segments without any labeled data. These tasks are particularly interesting in an unsupervised setting due to the difficulty and cost of obtaining dense image annotations, but existing unsupervised approaches struggle with complex scenes containing multiple objects. Differently from existing methods, which are purely based on deep learning, we take inspiration from traditional spectral segmentation methods by reframing image decomposition as a graph partitioning problem. Specifically, we examine the eigenvectors of the Laplacian of a feature affinity matrix from self-supervised networks. We find that these eigenvectors already decompose an image into meaningful segments, and can be readily used to localize objects in a scene. Furthermore, by clustering the features associated with these segments across a dataset, we can obtain well-delineated, nameable regions, i.e. semantic segmentations. Experiments on complex datasets (Pascal VOC, MS-COCO) demonstrate that our simple spectral method outperforms the state-of-the-art in unsupervised localization and segmentation by a significant margin. Furthermore, our method can be readily used for a variety of complex image editing tasks, such as background removal and compositing.

Stitched images provide a wide field-of-view (FoV) but suffer from unpleasant irregular boundaries. To deal with this problem, existing image rectangling methods devote to searching an initial mesh and optimizing a target mesh to form the mesh deformation in two stages. Then rectangular images can be generated by warping stitched images. However, these solutions only work for images with rich linear structures, leading to noticeable distortions for portraits and landscapes with non-linear objects. In this paper, we address these issues by proposing the first deep learning solution to image rectangling. Concretely, we predefine a rigid target mesh and only estimate an initial mesh to form the mesh deformation, contributing to a compact one-stage solution. The initial mesh is predicted using a fully convolutional network with a residual progressive regression strategy. To obtain results with high content fidelity, a comprehensive objective function is proposed to simultaneously encourage the boundary rectangular, mesh shape-preserving, and content perceptually natural. Besides, we build the first image stitching rectangling dataset with a large diversity in irregular boundaries and scenes. Extensive experiments demonstrate our superiority over traditional methods both quantitatively and qualitatively. The codes and dataset will be available.

Advection-diffusion equations describe a large family of natural transport processes, e.g., fluid flow, heat transfer, and wind transport. They are also used for optical flow and perfusion imaging computations. We develop a machine learning model, D^2-SONATA, built upon a stochastic advection-diffusion equation, which predicts the velocity and diffusion fields that drive 2D/3D image time-series of transport. In particular, our proposed model incorporates a model of transport atypicality, which isolates abnormal differences between expected normal transport behavior and the observed transport. In a medical context such a normal-abnormal decomposition can be used, for example, to quantify pathologies. Specifically, our model identifies the advection and diffusion contributions from the transport time-series and simultaneously predicts an anomaly value field to provide a decomposition into normal and abnormal advection and diffusion behavior. To achieve improved estimation performance for the velocity and diffusion-tensor fields underlying the advection-diffusion process and for the estimation of the anomaly fields, we create a 2D/3D anomaly-encoded advection-diffusion simulator, which allows for supervised learning. We further apply our model on a brain perfusion dataset from ischemic stroke patients via transfer learning. Extensive comparisons demonstrate that our model successfully distinguishes stroke lesions (abnormal) from normal brain regions, while reconstructing the underlying velocity and diffusion tensor fields.

Image deblurring is an ill-posed problem with multiple plausible solutions for a given input image. However, most existing methods produce a deterministic estimate of the clean image and are trained to minimize pixel-level distortion. These metrics are known to be poorly correlated with human perception, and often lead to unrealistic reconstructions. We present an alternative framework for blind deblurring based on conditional diffusion models. Unlike existing techniques, we train a stochastic sampler that refines the output of a deterministic predictor and is capable of producing a diverse set of plausible reconstructions for a given input. This leads to a significant improvement in perceptual quality over existing state-of-the-art methods across multiple standard benchmarks. Our predict-and-refine approach also enables much more efficient sampling compared to typical diffusion models. Combined with a carefully tuned network architecture and inference procedure, our method is competitive in terms of distortion metrics such as PSNR. These results show clear benefits of our diffusion-based method for deblurring and challenge the widely used strategy of producing a single, deterministic reconstruction.

Many flagship smartphone cameras now use a dedicated neural image signal processor (ISP) to render noisy raw sensor images to the final processed output. Training nightmode ISP networks relies on large-scale datasets of image pairs with: (1) a noisy raw image captured with a short exposure and a high ISO gain; and (2) a ground truth low-noise raw image captured with a long exposure and low ISO that has been rendered through the ISP. Capturing such image pairs is tedious and time-consuming, requiring careful setup to ensure alignment between the image pairs. In addition, ground truth images are often prone to motion blur due to the long exposure. To address this problem, we propose a method that synthesizes nighttime images from daytime images. Daytime images are easy to capture, exhibit low-noise (even on smartphone cameras) and rarely suffer from motion blur. We outline a processing framework to convert daytime raw images to have the appearance of realistic nighttime raw images with different levels of noise. Our procedure allows us to easily produce aligned noisy and clean nighttime image pairs. We show the effectiveness of our synthesis framework by training neural ISPs for nightmode rendering. Furthermore, we demonstrate that using our synthetic nighttime images together with small amounts of real data (e.g., 5% to 10%) yields performance almost on par with training exclusively on real nighttime images. Our dataset and code are available at https://github.com/SamsungLabs/day-to-night.

Dataset distillation is the task of synthesizing a small dataset such that a model trained on the synthetic set will match the test accuracy of the model trained on the full dataset. The task is extremely challenging as it often involves backpropagating through the full training process or assuming the strong constraint that a single training step on distilled data can only imitate a single step on real data. In this paper, we propose a new formulation that optimizes our distilled data to guide networks to a similar state as those trained on real data across many training steps. Given a network, we train it for several iterations on our distilled data and optimize the distilled data with respect to the distance between the synthetically trained parameters and the parameters trained on real data. To efficiently obtain the initial and target network parameters for large-scale datasets, we pre-compute and store training trajectories of expert networks trained on the real dataset. Our method handily outperforms existing methods and also allows us to distill with higher-resolution visual data.

Imaging in low light is extremely challenging due to low photon counts. Using sensitive CMOS cameras, it is currently possible to take videos at night under moonlight (0.05-0.3 lux illumination). In this paper, we demonstrate photorealistic video under starlight (no moon present, <0.001 lux) for the first time. To enable this, we develop a GAN-tuned physics-based noise model to more accurately represent camera noise at the lowest light levels. Using this noise model, we train a video denoiser using a combination of simulated noisy video clips and real noisy still images. We capture a 5-10 fps video dataset with significant motion at approximately 0.6-0.7 millilux with no active illumination. Comparing against alternative methods, we achieve improved video quality at the lowest light levels, demonstrating photorealistic video denoising in starlight for the first time.

Multiple instance learning (MIL) has been increasingly used in the classification of histopathology whole slide images (WSIs). However, MIL approaches for this specific classification problem still face unique challenges, particularly those related to small sample cohorts. In these, there are limited number of WSI slides (bags), while the resolution of a single WSI is huge, which leads to a large number of patches (instances) cropped from this slide. To address this issue, we propose to virtually enlarge the number of bags by introducing the concept of pseudo-bags, on which a double-tier MIL framework is built to effectively use the intrinsic features. Besides, we also contribute to deriving the instance probability under the framework of attention-based MIL, and utilize the derivation to help construct and analyze the proposed framework. The proposed method outperforms other latest methods on the CAMELYON-16 by substantially large margins, and is also better in performance on the TCGA lung cancer dataset. The proposed framework is ready to be extended for wider MIL applications. The code is available at: https://github.com/hrzhang1123/DTFD-MIL

We propose the deep progressive image compression using trit-planes (DPICT) algorithm, which is the first learning-based codec supporting fine granular scalability (FGS). First, we transform an image into a latent tensor using an analysis network. Then, we represent the latent tensor in ternary digits (trits) and encode it into a compressed bitstream trit-plane by trit-plane in the decreasing order of significance. Moreover, within each trit-plane, we sort the trits according to their rate-distortion priorities and transmit more important information first. Since the compression network is less optimized for the cases of using fewer trit-planes, we develop a postprocessing network for refining reconstructed images at low rates. Experimental results show that DPICT outperforms conventional progressive codecs significantly, while enabling FGS transmission. Codes are available at https://github.com/jaehanlee-mcl/DPICT.

Despite an increased demand for valuable data, the privacy concerns associated with sensitive datasets present a barrier to data sharing. One may use differentially private generative models to generate synthetic data. Unfortunately, generators are typically restricted to generating images of low-resolutions due to the limitation of noisy gradients. Here, we propose DPGEN, a network model designed to synthesize high-resolution natural images while satisfying differential privacy. In particular, we propose an energy-guided network trained on sanitized data to indicate the direction of the true data distribution via Langevin Markov chain Monte Carlo (MCMC) sampling method. In contrast to the state-of-the-art methods that can process only low-resolution images (e.g., MNIST and Fashion-MNIST), DPGEN can generate differentially private synthetic images with resolutions up to 128*128 with superior visual quality and data utility. Our code is available at https://github.com/chiamuyu/DPGEN

We present in this paper a novel denoising training method to speedup DETR (DEtection TRansformer) training and offer a deepened understanding of the slow convergence issue of DETR-like methods. We show that the slow convergence results from the instability of bipartite graph matching which causes inconsistent optimization goals in early training stages. To address this issue, except for the Hungarian loss, our method additionally feeds ground-truth bounding boxes with noises into Transformer decoder and trains the model to reconstruct the original boxes, which effectively reduces the bipartite graph matching difficulty and leads to a faster convergence. Our method is universal and can be easily plugged into any DETR-like methods by adding dozens of lines of code to achieve a remarkable improvement. As a result, our DN-DETR results in a remarkable improvement (+1.9AP) under the same setting and achieves the best result (AP 43.4 and 48.6 with 12 and 50 epochs of training respectively) among DETR-like methods with ResNet-50 backbone. Our code will be released after the blind review.

DIVeR builds on the key ideas of NeRF and its variants -- density models and volume rendering -- to learn 3D object models that can be rendered realistically from small numbers of images. In contrast to all previous NeRF methods, DIVeR uses deterministic rather than stochastic estimates of the volume rendering integral. DIVeR's representation is a voxel based field of features. To compute the volume rendering integral, a ray is broken into intervals, one per voxel; components of the volume rendering integral are estimated from the features for each interval using an MLP, and the components are aggregated. As a result, DIVeR can render thin translucent structures that are missed by other integrators. Furthermore, DIVeR's representation has semantics that is relatively exposed compared to other such methods -- moving feature vectors around in the voxel space results in natural edits. Extensive qualitative and quantitative comparisons to current state-of-the-art methods show that DIVeR produces models that (1) render at or above state-of-the-art quality, (2) are very small without being baked, (3) render very fast without being baked, and (4) can be edited in natural ways.

To ease the burden of labeling, unsupervised domain adaptation (UDA) aims to transfer knowledge in previous and related labeled datasets (sources) to a new unlabeled dataset (target). Despite impressive progress, prior methods always need to access the raw source data and develop data-dependent alignment approaches to recognize the target samples in a transductive learning manner, which may raise privacy concerns from source individuals. Several recent studies resort to an alternative solution by exploiting the well-trained white-box model from the source domain, yet, it may still leak the raw data via generative adversarial learning. This paper studies a practical and interesting setting for UDA, where only black-box source models (i.e., only network predictions are available) are provided during adaptation in the target domain. To solve this problem, we propose a new two-step knowledge adaptation framework called DIstill and fine-tuNE (DINE). Taking into consideration the target data structure, DINE first distills the knowledge from the source predictor to a customized target model, then fine-tunes the distilled model to further fit the target domain. Besides, neural networks are not required to be identical across domains in DINE, even allowing effective adaptation on a low-resource device. Empirical results on three UDA scenarios (i.e., single-source, multi-source, and partial-set) confirm that DINE achieves highly competitive performance compared to state-of-the-art data-dependent approaches. Code is available at https://github.com/tim-learn/DINE/.

Synthesizing high-quality realistic images from text descriptions is a challenging task. Existing text-to-image Generative Adversarial Networks generally employ a stacked architecture as the backbone yet still remain three flaws. First, the stacked architecture introduces the entanglements between generators of different image scales. Second, existing studies prefer to apply and fix extra networks in adversarial learning for text-image semantic consistency, which limits the supervision capability of these networks. Third, the cross-modal attention-based text-image fusion that widely adopted by previous works is limited on several special image scales because of the computational cost. To these ends, we propose a simpler but more effective Deep Fusion Generative Adversarial Networks (DF-GAN). To be specific, we propose: (i) a novel one-stage text-to-image backbone that directly synthesizes high-resolution images without entanglements between different generators, (ii) a novel Target-Aware Discriminator composed of Matching-Aware Gradient Penalty and One-Way Output, which enhances the text-image semantic consistency without introducing extra networks, (iii) a novel deep text-image fusion block, which deepens the fusion process to make a full fusion between text and visual features. Compared with current state-of-the-art methods, our proposed DF-GAN is simpler but more efficient to synthesize realistic and text-matching images and achieves better performance on widely used datasets. Code is available at https://github.com/tobran/DF-GAN.

We present cross-view transformers, an efficient attention-based model for map-view semantic segmentation from multiple cameras. Our architecture implicitly learns a mapping from individual camera views into a canonical map-view representation using a camera-aware cross-view attention mechanism. Each camera uses positional embeddings that depend on its intrinsic and extrinsic calibration. These embeddings allow a transformer to learn the mapping across different views without ever explicitly modeling it geometrically. The architecture consists of a convolutional image encoder for each view and cross-view transformer layers to infer a map-view semantic segmentation. Our model is simple, easily parallelizable, and runs in real-time. The presented architecture performs at state-of-the-art on the nuScenes dataset, with 4x faster inference speeds. Code is available at https://github.com/bradyz/cross_view_transformers.

Semi-supervised action recognition is a challenging but important task due to the high cost of data annotation. A common approach to this problem is to assign unlabeled data with pseudo-labels, which are then used as additional supervision in training. Typically in recent work, the pseudo-labels are obtained by training a model on the labeled data, and then using confident predictions from the model to teach itself.In this work, we propose a more effective pseudo-labeling scheme, called Cross-Model Pseudo-Labeling (CMPL). Concretely, we introduce a lightweight auxiliary network in addition to the primary backbone, and ask them to predict pseudo-labels for each other. We observe that, due to their different structural biases, these two models tend to learn complementary representations from the same video clips. Each model can thus benefit from its counterpart by utilizing cross-model predictions as supervision. Experiments on different data partition protocols demonstrate the significant improvement of our framework over existing alternatives. For example, CMPL achieves 7.6% and 25.1% Top-1 accuracy on Kinetics-400 and UCF-101 using only the RGB modality and 1% labeled data, outperforming our baseline model, FixMatch, by 9.0% and 10.3%, respectively.