ICCV 2021 论文列表

Learning-based image denoising methods have been bounded to situations where well-aligned noisy and clean images are given, or samples are synthesized from predetermined noise models, e.g., Gaussian. While recent generative noise modeling methods aim to simulate the unknown distribution of real-world noise, several limitations still exist. In a practical scenario, a noise generator should learn to simulate the general and complex noise distribution without using paired noisy and clean images. However, since existing methods are constructed on the unrealistic assumption of real-world noise, they tend to generate implausible patterns and cannot express complicated noise maps. Therefore, we introduce a Clean-to-Noisy image generation framework, namely C2N, to imitate complex real-world noise without using any paired examples. We construct the noise generator in C2N accordingly with each component of real-world noise characteristics to express a wide range of noise accurately. Combined with our C2N, conventional denoising CNNs can be trained to outperform existing unsupervised methods on challenging real-world benchmarks by a large margin.

We introduce BuildingNet: (a) a large-scale dataset of 3D building models whose exteriors are consistently labeled, and (b) a graph neural network that labels building meshes by analyzing spatial and structural relations of their geometric primitives. To create our dataset, we used crowdsourcing combined with expert guidance, resulting in 513K annotated mesh primitives, grouped into 292K semantic part components across 2K building models. The dataset covers several building categories, such as houses, churches, skyscrapers, town halls, libraries, and castles. We include a benchmark for evaluating mesh and point cloud labeling. Buildings have more challenging structural complexity compared to objects in existing benchmarks (e.g., ShapeNet, PartNet), thus, we hope that our dataset can nurture the development of algorithms that are able to cope with such large-scale geometric data for both vision and graphics tasks e.g., 3D semantic segmentation, part-based generative models, correspondences, texturing, and analysis of point cloud data acquired from real-world buildings. Finally, we show that our mesh-based graph neural network significantly improves performance over several baselines for labeling 3D meshes. Our project page www.buildingnet.org includes our dataset and code.

Volumetric design is the first and critical step for professional building design, where architects not only depict the rough 3D geometry of the building but also specify the programs to form a 2D layout on each floor. Though 2D layout generation for a single story has been widely studied, there is no developed method for multi-story buildings. This paper focuses on volumetric design generation conditioned on an input program graph. Instead of outputting dense 3D voxels, we propose a new 3D representation named voxel graph that is both compact and expressive for building geometries. Our generator is a cross-modal graph neural network that uses a pointer mechanism to connect the input program graph and the output voxel graph, and the whole pipeline is trained using the adversarial framework. The generated designs are evaluated qualitatively by a user study and quantitatively using three metrics: quality, diversity, and connectivity accuracy. We show that our model generates realistic 3D volumetric designs and outperforms previous methods and baselines.

Most successful self-supervised learning methods are trained to align the representations of two independent views from the data. State-of-the-art methods in video are inspired by image techniques, where these two views are similarly extracted by cropping and augmenting the resulting crop. However, these methods miss a crucial element in the video domain: time. We introduce BraVe, a self-supervised learning framework for video. In BraVe, one of the views has access to a narrow temporal window of the video while the other view has a broad access to the video content. Our models learn to generalise from the narrow view to the general content of the video. Furthermore, BraVe processes the views with different backbones, enabling the use of alternative augmentations or modalities into the broad view such as optical flow, randomly convolved RGB frames, audio or their combinations. We demonstrate that BraVe achieves state-of-the-art results in self-supervised representation learning on standard video and audio classification benchmarks including UCF101, HMDB51, Kinetics, ESC-50 and AudioSet.

Recently, video deblurring has attracted considerable research attention, and several works suggest that events at high time rate can benefit deblurring. In this paper, we develop a principled framework D2Nets for video deblurring to exploit non-consecutively blurry frames, and propose a flexible event fusion module (EFM) to bridge the gap between event-driven and video deblurring. In D2Nets, we propose to first detect nearest sharp frames (NSFs) using a bidirectional LSTM detector, and then perform deblurring guided by NSFs. Furthermore, the proposed EFM is flexible to be incorporated into D2Nets, in which events can be leveraged to notably boost the deblurring performance. EFM can also be easily incorporated into existing deblurring networks, making event-driven deblurring task benefit from state-of-the-art deblurring methods. On synthetic and real-world blurry datasets, our methods achieve better results than competing methods, and EFM not only benefits D2Nets but also significantly improves the competing deblurring networks.

The image-to-image translation (I2IT) model takes a target label or a reference image as the input, and changes a source into the specified target domain. The two types of synthesis, either label- or reference-based, have substantial differences. Particularly, the label-based synthesis reflects the common characteristics of the target domain, and the reference-based shows the specific style similar to the reference. This paper intends to bridge the gap between them in the task of multi-attribute I2IT. We design the label- and reference-based encoding modules (LEM and REM) to compare the domain differences. They first transfer the source image and target label (or reference) into a common embedding space, by providing the opposite directions through the attribute difference vector. Then the two embeddings are simply fused together to form the latent code S_ rand (or S_ ref ), reflecting the domain style differences, which is injected into each layer of the generator by SPADE. To link LEM and REM, so that two types of results benefit each other, we encourage the two latent codes to be close, and set up the cycle consistency between the forward and backward translations on them. Moreover, the interpolation between the S_ rand and S_ ref is also used to synthesize an extra image. Experiments show that label- and reference-based synthesis are indeed mutually promoted, so that we can have the diverse results from LEM, and high quality results with the similar style of the reference.

Depth estimation is a long-lasting yet important task in computer vision. Most of the previous works try to estimate depth from input images and assume images are all-in-focus (AiF), which is less common in real-world applications. On the other hand, a few works take defocus blur into account and consider it as another cue for depth estimation. In this paper, we propose a method to estimate not only a depth map but an AiF image from a set of images with different focus positions (known as a focal stack). We design a shared architecture to exploit the relationship between depth and AiF estimation. As a result, the proposed method can be trained either supervisedly with ground truth depth, or unsupervisedly with AiF images as supervisory signals. We show in various experiments that our method outperforms the state-of-the-art methods both quantitatively and qualitatively, and also has higher efficiency in inference time.

Current 3D single object tracking approaches track the target based on a feature comparison between the target template and the search area. However, due to the common occlusion in LiDAR scans, it is non-trivial to conduct accurate feature comparisons on severe sparse and incomplete shapes. In this work, we exploit the ground truth bounding box given in the first frame as a strong cue to enhance the feature description of the target object, enabling a more accurate feature comparison in a simple yet effective way. In particular, we first propose the BoxCloud, an informative and robust representation, to depict an object using the point-to-box relation. We further design an efficient box-aware feature fusion module, which leverages the aforementioned BoxCloud for reliable feature matching and embedding. Integrating the proposed general components into an existing model P2B, we construct a superior box-aware tracker (BAT). Experiments confirm that our proposed BAT outperforms the previous state-of-the-art by a large margin on both KITTI and NuScenes benchmarks, achieving a 12.8% improvement in terms of precision while running 20% faster.

Many video analysis tasks require temporal localization for the detection of content changes. However, most existing models developed for these tasks are pre-trained on general video action classification tasks. This is due to large scale annotation of temporal boundaries in untrimmed videos being expensive. Therefore, no suitable datasets exist that enable pre-training in a manner sensitive to temporal boundaries. In this paper for the first time, we investigate model pre-training for temporal localization by introducing a novel boundary-sensitive pretext (BSP) task. Instead of relying on costly manual annotations of temporal boundaries, we propose to synthesize temporal boundaries in existing video action classification datasets. By defining different ways of synthesizing boundaries, BSP can then be simply conducted in a self-supervised manner via the classification of the boundary types. This enables the learning of video representations that are much more transferable to downstream temporal localization tasks. Extensive experiments show that the proposed BSP is superior and complementary to the existing action classification-based pre-training counterpart, and achieves new state-of-the-art performance on several temporal localization tasks. Please visit our website for more details https://frostinassiky.github.io/bsp.

A myriad of recent breakthroughs in hand-crafted neural architectures for visual recognition have highlighted the urgent need to explore hybrid architectures consisting of diversified building blocks. Meanwhile, neural architecture search methods are surging with an expectation to reduce human efforts. However, whether NAS methods can efficiently and effectively handle diversified search spaces with disparate candidates (e.g. CNNs and transformers) is still an open question. In this work, we present Block-wisely Self-supervised Neural Architecture Search (BossNAS), an unsupervised NAS method that addresses the problem of inaccurate architecture rating caused by large weight-sharing space and biased supervision in previous methods. More specifically, we factorize the search space into blocks and utilize a novel self-supervised training scheme, named ensemble bootstrapping, to train each block separately before searching them as a whole towards the population center. Additionally, we present HyTra search space, a fabric-like hybrid CNN-transformer search space with searchable down-sampling positions. On this challenging search space, our searched model, BossNet-T, achieves up to 82.5% accuracy on ImageNet, surpassing EfficientNet by 2.4% with comparable compute time. Moreover, our method achieves superior architecture rating accuracy with 0.78 and 0.76 Spearman correlation on the canonical MBConv search space with ImageNet and on NATS-Bench size search space with CIFAR-100, respectively, surpassing state-of-the-art NAS methods.

Geometric feature extraction is a crucial component of point cloud registration pipelines. Recent work has demonstrated how supervised learning can be leveraged to learn better and more compact 3D features. However, those approaches' reliance on ground-truth annotation limits their scalability. We propose BYOC: a self-supervised approach that learns visual and geometric features from RGB-D video without relying on ground-truth pose or correspondence. Our key observation is that randomly-initialized CNNs readily provide us with good correspondences; allowing us to bootstrap the learning of both visual and geometric features. Our approach combines classic ideas from point cloud registration with more recent representation learning approaches. We evaluate our approach on indoor scene datasets and find that our method outperforms traditional and learned descriptors, while being competitive with current state-of-the-art supervised approaches.

State of the art (SOTA) few-shot learning (FSL) methods suffer significant performance drop in the presence of domain differences between source and target datasets. The strong discrimination ability on the source dataset does not necessarily translate to high classification accuracy on the target dataset. In this work, we address this cross-domain few-shot learning (CDFSL) problem by boosting the generalization capability of the model. Specifically, we teach the model to capture broader variations of the feature distributions with a novel noise-enhanced supervised autoencoder (NSAE). NSAE trains the model by jointly reconstructing inputs and predicting the labels of inputs as well as their reconstructed pairs. Theoretical analysis based on intra-class correlation (ICC) shows that the feature embeddings learned from NSAE have stronger discrimination and generalization abilities in the target domain. We also take advantage of NSAE structure and propose a two-step fine-tuning procedure that achieves better adaption and improves classification performance in the target domain. Extensive experiments and ablation studies are conducted to demonstrate the effectiveness of the proposed method. Experimental results show that our proposed method consistently outperforms SOTA methods under various conditions.

Weakly-supervised object detection (WSOD) has emerged as an inspiring recent topic to avoid expensive instance-level object annotations. However, the bounding boxes of most existing WSOD methods are mainly determined by precomputed proposals, thereby being limited in precise object localization. In this paper, we defend the problem setting for improving localization performance by leveraging the bounding box regression knowledge from a well-annotated auxiliary dataset. First, we use the well-annotated auxiliary dataset to explore a series of learnable bounding box adjusters (LBBAs) in a multi-stage training manner, which is class-agnostic. Then, only LBBAs and a weakly-annotated dataset with non-overlapped classes are used for training LBBA-boosted WSOD. As such, our LBBAs are practically more convenient and economical to implement while avoiding the leakage of the auxiliary well-annotated dataset. In particular, we formulate learning bounding box adjusters as a bi-level optimization problem and suggest an EM-like multi-stage training algorithm. Then, a multi-stage scheme is further presented for LBBA-boosted WSOD. Additionally, a masking strategy is adopted to improve proposal classification. Experimental results verify the effectiveness of our method. Our method performs favorably against state-of-the-art WSOD methods and knowledge transfer model with similar problem setting. Code is publicly available at https://github.com/DongSky/lbba_boosted_wsod.

In this paper, we propose enhancing monocular depth estimation by adding 3D points as depth guidance. Unlike existing depth completion methods, our approach performs well on extremely sparse and unevenly distributed point clouds, which makes it agnostic to the source of the 3D points. We achieve this by introducing a novel multi-scale 3D point fusion network that is both lightweight and efficient. We demonstrate its versatility on two different depth estimation problems where the 3D points have been acquired with conventional structure-from-motion and LiDAR. In both cases, our network performs on par with state-of-the-art depth completion methods and achieves significantly higher accuracy when only a small number of points is used while being more compact in terms of the number of parameters. We show that our method outperforms some contemporary deep learning based multi-view stereo and structure-from-motion methods both in accuracy and in compactness.

Detecting pedestrians and their associated faces jointly is a challenging task.On one hand, body or face could be absent because of occlusion or non-frontal human pose.On the other hand, the association becomes difficult or even miss-leading in crowded scenes due to the lack of strong correlational evidence. This paper proposes Body-Face Joint (BFJ) detector, a novel framework for detecting bodies and their faces with accurate correspondance. We follow the classical multi-class detector design by detecting body and face in parallel but with two key contributions. First, we propose an Embedding Matching Loss (EML) to learn an associative embedding for matching body and face of the same person. Second, we introduce a novel concept, "head hook", to bridge the gap of matching body and faces spatially. With the new semantical and geometrical sources of information, BFJ greatly reduces the difficulty of detecting body and face in pairs. Since the problem is unexplored yet, we design a new metric named log-average miss matching rate (mMR^ -2 ) to evaluate the association performance and extend the CrowdHuman and CityPersons benchmarks by annotating each face box. Experiments show that our BFJ detector can maintain state-of-the-art performance in pedestrian detection on both one-stage and two-stage structures while greatly outperform various body-face association strategies. Code and datasets will be released soon.

City modeling is the foundation for computational urban planning, navigation, and entertainment. In this work, we present the first generative model of city blocks named BlockPlanner, and showcase its ability to synthesize valid city blocks with varying land lots configurations. We propose a novel vectorized city block representation utilizing a ring topology and a two-tier graph to capture the global and local structures of a city block. Each land lot is abstracted into a vector representation covering both its 3D geometry and land use semantics. Such vectorized representation enables us to deploy a lightweight network to capture the underlying distribution of land lots configuration in a city block. To enforce intrinsic spatial constraints of a valid city block, a set of effective loss functions are imposed to shape rational results. We contribute a pilot city block dataset to demonstrate the effectiveness and efficiency of our representation and framework over the state-of-the-art. Notably, our BlockPlanner is also able to edit and manipulate city blocks, enabling several useful applications, e.g., topology refinement and footprint generation.

In this paper we propose BlockCopy, a scheme that accelerates pretrained frame-based CNNs to process video more efficiently, compared to standard frame-by-frame processing. To this end, a lightweight policy network determines important regions in an image, and operations are applied on selected regions only, using custom block-sparse convolutions. Features of non-selected regions are simply copied from the preceding frame, reducing the number of computations and latency. The execution policy is trained using reinforcement learning in an online fashion without requiring ground truth annotations. Our universal framework is demonstrated on dense prediction tasks such as pedestrian detection, instance segmentation and semantic segmentation, using both state of the art (Center and Scale Predictor, MGAN, SwiftNet) and standard baseline networks (Mask-RCNN, DeepLabV3+). BlockCopy achieves significant FLOPS savings and inference speedup with minimal impact on accuracy.

Although deep neural networks (DNNs) have made rapid progress in recent years, they are vulnerable in adversarial environments. A malicious backdoor could be embedded in a model by poisoning the training dataset, whose intention is to make the infected model give wrong predictions during inference when the specific trigger appears. To mitigate the potential threats of backdoor attacks, various backdoor detection and defense methods have been proposed. However, the existing techniques usually require the poisoned training data or access to the white-box model, which is commonly unavailable in practice. In this paper, we propose a black-box backdoor detection (B3D) method to identify backdoor attacks with only query access to the model. We introduce a gradient-free optimization algorithm to reverse-engineer the potential trigger for each class, which helps to reveal the existence of backdoor attacks. In addition to backdoor detection, we also propose a simple strategy for reliable predictions using the identified backdoored models. Extensive experiments on hundreds of DNN models trained on several datasets corroborate the effectiveness of our method under the black-box setting against various backdoor attacks.

Mixed-precision networks allow for a variable bit-width quantization for every layer in the network. A major limitation of existing work is that the bit-width for each layer must be predefined during training time. This allows little flexibility if the characteristics of the device on which the network is deployed change during runtime. In this work, we propose Bit-Mixer, the very first method to train a meta-quantized network where during test time any layer can change its bit-width without affecting at all the overall network's ability for highly accurate inference. To this end, we make 2 key contributions: (a) Transitional Batch-Norms, and (b) a 3-stage optimization process which is shown capable of training such a network. We show that our method can result in mixed precision networks that exhibit the desirable flexibility properties for on-device deployment without compromising accuracy. Code will be made available.