CVPR 2021 论文列表

In the animation industry, cartoon videos are usually produced at low frame rate since hand drawing of such frames is costly and time-consuming. Therefore, it is desirable to develop computational models that can automatically interpolate the in-between animation frames. However, existing video interpolation methods fail to produce satisfying results on animation data. Compared to natural videos, animation videos possess two unique characteristics that make frame interpolation difficult: 1) cartoons comprise lines and smooth color pieces. The smooth areas lack textures and make it difficult to estimate accurate motions on animation videos. 2) cartoons express stories via exaggeration. Some of the motions are non-linear and extremely large. In this work, we formally define and study the animation video interpolation problem for the first time. To address the aforementioned challenges, we propose an effective framework, AnimeInterp, with two dedicated modules in a coarse-to-fine manner. Specifically, 1) Segment-Guided Matching resolves the "lack of textures" challenge by exploiting global matching among color pieces that are piece-wise coherent. 2) Recurrent Flow Refinement resolves the "non-linear and extremely large motion" challenge by recurrent predictions using a transformer-like architecture. To facilitate comprehensive training and evaluations, we build a large-scale animation triplet dataset, ATD-12K, which comprises 12,000 triplets with rich annotations. Extensive experiments demonstrate that our approach outperforms existing state-of-the-art interpolation methods for animation videos. Notably, AnimeInterp shows favorable perceptual quality and robustness for animation scenarios in the wild. The proposed dataset and code are available at https://github.com/lisiyao21/AnimeInterp/.

In recent years, a number of approaches based on 2D or 3D convolutional neural networks (CNN) have emerged for video action recognition, achieving state-of-the-art results on several large-scale benchmark datasets. In this paper, we carry out in-depth comparative analysis to better understand the differences between these approaches and the progress made by them. To this end, we develop an unified framework for both 2D-CNN and 3D-CNN action models, which enables us to remove bells and whistles and provides a common ground for fair comparison. We then conduct an effort towards a large-scale analysis involving over 300 action recognition models. Our comprehensive analysis reveals that a) a significant leap is made in efficiency for action recognition, but not in accuracy; b) 2D-CNN and 3D-CNN models behave similarly in terms of spatio-temporal representation abilities and transferability. Our codes are available at https://github.com/IBM/action-recognition-pytorch.

Active Surface Models have a long history of being useful to model complex 3D surfaces. But only Active Contours have been used in conjunction with deep networks, and then only to produce the data term as well as meta-parameter maps controlling them. In this paper, we advocate a much tighter integration. We introduce layers that implement them that can be integrated seamlessly into Graph Convolutional Networks to enforce sophisticated smoothness priors at an acceptable computational cost. We will show that the resulting Deep Active Surface Models outperform equivalent architectures that use traditional regularization loss terms to impose smoothness priors for 3D surface reconstruction from 2D images and for 3D volume segmentation.

Convolution is one of the basic building blocks of CNN architectures. Despite its common use, standard convolution has two main shortcomings: Content-agnostic and Computation-heavy. Dynamic filters are content-adaptive, while further increasing the computational overhead. Depth-wise convolution is a lightweight variant, but it usually leads to a drop in CNN performance or requires a larger number of channels. In this work, we propose the Decoupled Dynamic Filter (DDF) that can simultaneously tackle both of these shortcomings. Inspired by recent advances in attention, DDF decouples a depth-wise dynamic filter into spatial and channel dynamic filters. This decomposition considerably reduces the number of parameters and limits computational costs to the same level as depth-wise convolution. Meanwhile, we observe a significant boost in performance when replacing standard convolution with DDF in classification networks. ResNet50 / 101 get improved by 1.9% and 1.3% on the top-1 accuracy, while their computational costs are reduced by nearly half. Experiments on the detection and joint upsampling networks also demonstrate the superior performance of the DDF upsampling variant (DDF-Up) in comparison with standard convolution and specialized content-adaptive layers.

In real-world image enhancement, it is often challenging (if not impossible) to acquire ground-truth data, preventing the adoption of distance metrics for objective quality assessment. As a result, one often resorts to subjective quality assessment, the most straightforward and reliable means of evaluating image enhancement. Conventional subjective testing requires manually pre-selecting a small set of visual examples, which may suffer from three sources of biases: 1) sampling bias due to the extremely sparse distribution of the selected samples in the image space; 2) algorithmic bias due to potential overfitting the selected samples; 3) subjective bias due to further potential cherry-picking test results. This eventually makes the field of real-world image enhancement more of an art than a science. Here we take steps towards debiasing conventional subjective assessment by automatically sampling a set of adaptive and diverse images for subsequent testing. This is achieved by casting sample selection into a joint maximization of the discrepancy between the enhancers and the diversity among the selected input images. Careful visual inspection on the resulting enhanced images provides a debiased ranking of the enhancement algorithms. We demonstrate our subjective assessment method using three popular and practically demanding image enhancement tasks: dehazing, super-resolution, and low-light enhancement.

With the advent of Neural Radiance Fields (NeRF), neural networks can now render novel views of a 3D scene with quality that fools the human eye. Yet, generating these images is very computationally intensive, limiting their applicability in practical scenarios. In this paper, we propose a technique based on spatial decomposition capable of mitigating this issue. Our key observation is that there are diminishing returns in employing larger (deeper and/or wider) networks. Hence, we propose to spatially decompose a scene and dedicate smaller networks for each decomposed part. When working together, these networks can render the whole scene. This allows us near-constant inference time regardless of the number of decomposed parts. Moreover, we show that a Voronoi spatial decomposition is preferable for this purpose, as it is provably compatible with the Painter's Algorithm for efficient and GPU-friendly rendering. Our experiments show that for real-world scenes, our method provides up to 3x more efficient inference than NeRF (with the same rendering quality), or an improvement of up to 1.0 dB in PSNR (for the same inference cost).

The difficulty of obtaining paired data remains a major bottleneck for learning image restoration and enhancement models for real-world applications. Current strategies aim to synthesize realistic training data by modeling noise and degradations that appear in real-world settings. We propose DeFlow, a method for learning stochastic image degradations from unpaired data. Our approach is based on a novel unpaired learning formulation for conditional normalizing flows. We model the degradation process in the latent space of a shared flow encoder-decoder network. This allows us to learn the conditional distribution of a noisy image given the clean input by solely minimizing the negative log-likelihood of the marginal distributions. We validate our DeFlow formulation on the task of joint image restoration and super-resolution. The models trained with the synthetic data generated by DeFlow outperform previous learnable approaches on three recent datasets. Code and trained models will be made available at: https://github.com/volflow/DeFlow

Objects moving at high speed appear significantly blurred when captured with cameras. The blurry appearance is especially ambiguous when the object has complex shape or texture. In such cases, classical methods, or even humans, are unable to recover the object's appearance and motion. We propose a method that, given a single image with its estimated background, outputs the object's appearance and position in a series of sub-frames as if captured by a high-speed camera (i.e. temporal super-resolution). The proposed generative model embeds an image of the blurred object into a latent space representation, disentangles the background, and renders the sharp appearance. Inspired by the image formation model, we design novel self-supervised loss function terms that boost performance and show good generalization capabilities. The proposed DeFMO method is trained on a complex synthetic dataset, yet it performs well on real-world data from several datasets. DeFMO outperforms the state of the art and generates high-quality temporal super-resolution frames.

User-intended visual content fills the hole regions of an input image in the image editing scenario. The coarse lowlevel inputs, which typically consist of sparse sketch lines and color dots, convey user intentions for content creation (i.e., free-form editing). While existing methods combine an input image and these low-level controls for CNN inputs, the corresponding feature representations are not sufficient to convey user intentions, leading to unfaithfully generated content. In this paper, we propose DeFLOCNet which is based on a deep encoder-decoder CNN to retain the guidance of these controls in the deep feature representations. In each skip connection layer, we design a structure generation block. Instead of attaching low-level controls to an input image, we inject these controls directly into each structure generation block for sketch line refinement and color propagation in the CNN feature space. We then concatenate the modulated features with the original decoder features for structure generation. Meanwhile, DeFLOCNet involves another decoder branch for texture generation and detail enhancement. Both structures and textures are rendered in the decoder, leading to user-intended editing results. Experiments on benchmarks indicate that DeFLOCNet effectively transforms different user intentions to create visually pleasing content.

Recent works have shown exciting results in unsupervised image de-rendering--learning to decompose 3D shape, appearance, and lighting from single-image collections without explicit supervision. However, many of these assume simplistic material and lighting models. We propose a method, termed RADAR, that can recover environment illumination and surface materials from real single-image collections, relying neither on explicit 3D supervision, nor on multi-view or multi-light images. Specifically, we focus on rotationally symmetric artefacts that exhibit challenging surface properties including specular reflections, such as vases. We introduce a novel self-supervised albedo discriminator, which allows the model to recover plausible albedo without requiring any ground-truth during training. In conjunction with a shape reconstruction module exploiting rotational symmetry, we present an end-to-end learning framework that is able to de-render the world's revolutionary artefacts. We conduct experiments on a real vase dataset and demonstrate compelling decomposition results, allowing for applications including free-viewpoint rendering and relighting. More results and code at: https://sorderender.github.io/.

We introduce DatasetGAN: an automatic procedure to generate massive datasets of high-quality semantically segmented images requiring minimal human effort. Current deep networks are extremely data-hungry, benefiting from training on large-scale datasets, which are time-consuming to annotate. Our method relies on the power of recent GANs to generate realistic images. We show how the GAN latent code can be decoded to produce a semantic segmentation of the image. Training the decoder only needs a few labeled examples to generalize to the rest of the latent space, resulting in an infinite annotated dataset generator! These generated datasets can then be used for training any computer vision architecture just as real datasets are. As only a few images need to be manually segmented, it becomes possible to annotate images in extreme detail and generate datasets with rich object and part segmentations. To showcase the power of our approach, we generated datasets for 7 image segmentation tasks which include pixel-level labels for 34 human face parts, and 32 car parts. Our approach outperforms all semi-supervised baselines significantly and is on par with fully supervised methods using labor-intensive annotations.

In this paper, we delve into semi-supervised object detection where unlabeled images are leveraged to break through the upper bound of fully-supervised object detection models. Previous semi-supervised methods based on pseudo labels are severely degenerated by noise and prone to overfit to noisy labels, thus are deficient in learning different unlabeled knowledge well. To address this issue, we propose a data-uncertainty guided multi-phase learning method for semi-supervised object detection. We comprehensively consider divergent types of unlabeled images according to their difficulty levels, utilize them in different phases and ensemble models from different phases together to generate ultimate results. Image uncertainty guided easy data selection and region uncertainty guided RoI Re-weighting are involved in multi-phase learning and enable the detector to concentrate on more certain knowledge. Through extensive experiments on PASCAL VOC and MS COCO, we demonstrate that our method behaves extraordinarily compared to baseline approaches and outperforms them by a large margin, more than 3% on VOC and 2% on COCO.

Current model extraction attacks assume that the adversary has access to a surrogate dataset with characteristics similar to the proprietary data used to train the victim model. This requirement precludes the use of existing model extraction techniques on valuable models, such as those trained on rare or hard to acquire datasets. In contrast, we propose data-free model extraction methods that do not require a surrogate dataset. Our approach adapts techniques from the area of data-free knowledge transfer for model extraction. As part of our study, we identify that the choice of loss is critical to ensuring that the extracted model is an accurate replica of the victim model. Furthermore, we address difficulties arising from the adversary's limited access to the victim model in a black-box setting. For example, we recover the model's logits from its probability predictions to approximate gradients. We find that the proposed data-free model extraction approach achieves high-accuracy with reasonable query complexity -- 0.99x and 0.92x the victim model accuracy on SVHN and CIFAR-10 datasets given 2M and 20M queries respectively.

Convolutional network compression methods require training data for achieving acceptable results, but training data is routinely unavailable due to some privacy and transmission limitations. Therefore, recent works focus on learning efficient networks without original training data, i.e., data-free model compression. Wherein, most of existing algorithms are developed for image recognition or segmentation tasks. In this paper, we study the data-free compression approach for single image super-resolution (SISR) task which is widely used in mobile phones and smart cameras. Specifically, we analyze the relationship between the outputs and inputs from the pre-trained network and explore a generator with a series of loss functions for maximally capturing useful information. The generator is then trained for synthesizing training samples which have similar distribution to that of the original data. To further alleviate the training difficulty of the student network using only the synthetic data, we introduce a progressive distillation scheme. Experiments on various datasets and architectures demonstrate that the proposed method is able to be utilized for effectively learning portable student networks without the original data, e.g., with 0.16dB PSNR drop on Set5 for x2 super resolution. Code will be available at https://github.com/huaweinoah/Data-Efficient-Model-Compression.

In deep learning applications, the architectures of deep neural networks are crucial in achieving high accuracy. Many methods have been proposed to search for high-performance neural architectures automatically. However, these searched architectures are prone to adversarial attacks. A small perturbation of the input data can render the architecture to change prediction outcomes significantly. To address this problem, we propose methods to perform differentiable searches of robust neural architectures. In our methods, two differentiable metrics are defined to measure architectures' robustness, based on certified lower bound and Jacobian norm bound. Then we search for robust architectures by maximizing the robustness metrics. Different from previous approaches which aim to improve architectures' robustness in an implicit way: performing adversarial training and injecting random noise, our methods explicitly and directly maximize robustness metrics to harvest robust architectures. On CIFAR-10, ImageNet, and MNIST, we perform game-based evaluation and verification-based evaluation on the robustness of our methods. The experimental results show that our methods 1) are more robust to various norm-bound attacks than several robust NAS baselines; 2) are more accurate than baselines when there are no attacks; 3) have significantly higher certified lower bounds than baselines.

Compared to 2D object bounding-box labeling, it is very difficult for humans to annotate 3D object poses, especially when depth images of scenes are unavailable. This paper investigates whether we can estimate the object poses effectively when only RGB images and 2D object annotations are given. To this end, we present a two-step pose estimation framework to attain 6DoF object poses from 2D object bounding-boxes. In the first step, the framework learns to segment objects from real and synthetic data in a weakly-supervised fashion, and the segmentation masks will act as a prior for pose estimation. In the second step, we design a dual-scale pose estimation network, namely DSC-PoseNet, to predict object poses by employing a differential renderer. To be specific, our DSC-PoseNet firstly predicts object poses in the original image scale by comparing the segmentation masks and the rendered visible object masks. Then, we resize object regions to a fixed scale to estimate poses once again. In this fashion, we eliminate large scale variations and focus on rotation estimation, thus facilitating pose estimation. Moreover, we exploit the initial pose estimation to generate pseudo ground-truth to train our DSC-PoseNet in a self-supervised manner. The estimation results in these two scales are ensembled as our final pose estimation. Extensive experiments on widely-used benchmarks demonstrate that our method outperforms state-of-the-art models trained on synthetic data by a large margin and even is on par with several fully-supervised methods.

In this paper, we present DRANet, a network architecture that disentangles image representations and transfers the visual attributes in a latent space for unsupervised cross-domain adaptation. Unlike the existing domain adaptation methods that learn associated features sharing a domain, DRANet preserves the distinctiveness of each domain's characteristics. Our model encodes individual representations of content (scene structure) and style (artistic appearance) from both source and target images. Then, it adapts the domain by incorporating the transferred style factor into the content factor along with learnable weights specified for each domain. This learning framework allows bi-/multi-directional domain adaptation with a single encoder-decoder network and aligns their domain shift. Additionally, we propose a content-adaptive domain transfer module that helps retain scene structure while transferring style. Extensive experiments show our model successfully separates content-style factors and synthesizes visually pleasing domain-transferred images. The proposed method demonstrates state-of-the-art performance on standard digit classification tasks as well as semantic segmentation tasks.

Differentiable Architecture Search (DARTS) has attracted extensive attention due to its efficiency in searching for cell structures. DARTS mainly focuses on the operation search and derives the cell topology from the operation weights. However, the operation weights can not indicate the importance of cell topology and result in poor topology rating correctness. To tackle this, we propose to Decouple the Operation and Topology Search (DOTS), which decouples the topology representation from operation weights and makes an explicit topology search. DOTS is achieved by introducing a topology search space that contains combinations of candidate edges. The proposed search space directly reflects the search objective and can be easily extended to support a flexible number of edges in the searched cell. Existing gradient-based NAS methods can be incorporated into DOTS for further improvement by the topology search. Considering that some operations (e.g., Skip-Connection) can affect the topology, we propose a group operation search scheme to preserve topology-related operations for a better topology search. The experiments on CIFAR10/100 and ImageNet demonstrate that DOTS is an effective solution for differentiable NAS. The code is released at https://github.com/guyuchao/DOTS.

Recent deep learning models have shown remarkable performance in image classification. While these deep learning systems are getting closer to practical deployment, the common assumption made about data is that it does not carry any sensitive information. This assumption may not hold for many practical cases, especially in the domain where an individual's personal information is involved, like healthcare and facial recognition systems. We posit that selectively removing features in this latent space can protect the sensitive information and provide better privacy-utility trade-off. Consequently, we propose DISCO which learns a dynamic and data driven pruning filter to selectively obfuscate sensitive information in the feature space. We propose diverse attack schemes for sensitive inputs and attributes and demonstrate the effectiveness of DISCO against state-of-the-art methods through quantitative and qualitative evaluation. Finally, we also release an evaluation benchmark dataset of 1 million sensitive representations to encourage rigorous exploration of novel attack and defense schemes at https://github.com/splitlearning/InferenceBenchmark