CVPR 2021 Papers

Font generation is a challenging problem especially for some writing systems that consist of a large number of characters and has attracted a lot of attention in recent years. However, existing methods for font generation are often in supervised learning. They require a large number of paired data, which is labor-intensive and expensive to collect. Besides, common image-to-image translation models often define style as the set of textures and colors, which cannot be directly applied to font generation. To address these problems, we propose novel deformable generative networks for unsupervised font generation (DG-Font). We introduce a feature deformation skip connection (FDSC) which predicts pairs of displacement maps and employs the predicted maps to apply deformable convolution to the low-level feature maps from the content encoder. The outputs of FDSC are fed into a mixer to generate the final results. Taking advantage of FDSC, the mixer outputs a high-quality character with a complete structure. To further improve the quality of generated images, we use three deformable convolution layers in the content encoder to learn style-invariant feature representations. Experiments demonstrate that our model generates characters in higher quality than state-of-art methods. The source code is available at https://github.com/ecnuycxie/DG-Font.

We address the problem of class incremental learning, which is a core step towards achieving adaptive vision intelligence. In particular, we consider the task setting of incremental learning with limited memory and aim to achieve a better stability-plasticity trade-off. To this end, we propose a novel two-stage learning approach that utilizes a dynamically expandable representation for more effective incremental concept modeling. Specifically, at each incremental step, we freeze the previously learned representation and augment it with additional feature dimensions from a new learnable feature extractor. Moreover, we dynamically expand the representation according to the complexity of novel concepts by introducing a channel-level mask-based pruning strategy. This enables us to integrate new visual concepts with retaining learned knowledge. Furthermore, we introduce an auxiliary loss to encourage the model to learn diverse and discriminate features for novel concepts. We conduct extensive experiments on the three class incremental learning benchmarks and our method consistently outperforms other methods with a large margin.

We introduce a deep generative network for 3D shape detailization, akin to stylization with the style being geometric details. We address the challenge of creating large varieties of high-resolution and detailed 3D geometry from a small set of exemplars by treating the problem as that of geometric detail transfer. Given a low-resolution coarse voxel shape, our network refines it, via voxel upsampling, into a higher-resolution shape enriched with geometric details. The output shape preserves the overall structure (or content) of the input, while its detail generation is conditioned on an input "style code" corresponding to a detailed exemplar. Our 3D detailization via conditional refinement is realized by a generative adversarial network, coined DECOR-GAN. The network utilizes a 3D CNN generator for upsampling coarse voxels and a 3D PatchGAN discriminator to enforce local patches of the generated model to be similar to those in the training detailed shapes. During testing, a style code is fed into the generator to condition the refinement. We demonstrate that our method can refine a coarse shape into a variety of detailed shapes with different styles. The generated results are evaluated in terms of content preservation, plausibility, and diversity. Comprehensive ablation studies are conducted to validate our network designs. Code is available at https://github.com/czq142857/DECOR-GAN.

Binary grid mask representation is broadly used in instance segmentation. A representative instantiation is Mask R-CNN which predicts masks on a 28*28 binary grid. Generally, a low-resolution grid is not sufficient to capture the details, while a high-resolution grid dramatically increases the training complexity. In this paper, we propose a new mask representation by applying the discrete cosine transform(DCT) to encode the high-resolution binary grid mask into a compact vector. Our method, termed DCT-Mask, could be easily integrated into most pixel-based instance segmentation methods. Without any bells and whistles, DCT-Mask yields significant gains on different frameworks, backbones, datasets, and training schedules. It does not require any pre-processing or pre-training, and almost no harm to the running speed. Especially, for higher-quality annotations and more complex backbones, our method has a greater improvement. Moreover, we analyze the performance of our method from the perspective of the quality of mask representation. The main reason why DCT-Mask works well is that it obtains a high-quality mask representation with low complexity.

Existing NAS methods for dense image prediction tasks usually compromise on restricted search space or search on proxy task to meet the achievable computational demands. To allow as wide as possible network architectures and avoid the gap between realistic and proxy setting, we propose a novel Densely Connected NAS (DCNAS) framework, which directly searches the optimal network structures for the multi-scale representations of visual information, over a large-scale target dataset without proxy. Specifically, by connecting cells with each other using learnable weights, we introduce a densely connected search space to cover an abundance of mainstream network designs. Moreover, by combining both path-level and channel-level sampling strategies, we design a fusion module and mixture layer to reduce the memory consumption of ample search space, hence favoring the proxyless searching. Compared with contemporary works, experiments reveal that the proxyless searching scheme is capable of bridging the gap between searching and training environments. Further, DCNAS achieves new state-of-the-art performances on public semantic image segmentation benchmarks, including 84.3% on Cityscapes, and 86.9% on PASCAL VOC 2012. We also retain leading performances when evaluating the architecture on the more challenging ADE20K and PASCAL-Context dataset.

In real application scenarios, the performance of deep networks may be degraded when the dataset contains noisy labels. Existing methods for learning with noisy labels are limited by two aspects. Firstly, methods based on the noise probability modeling can only be applied to class-level noisy labels. Secondly, others based on the memorization effect outperform in synthetic noise but get weak promotion in real-world noisy datasets. To solve these problems, this paper proposes a novel label-noise robust method named Discrepant Adversarial Training (DAT). The DAT method has ability of enforcing prominent feature extraction by matching feature distribution between clean and noisy data. Therefore, under the noise-free feature representation, the deep network can simply output the correct result. To better capture the divergence between the noisy and clean distribution, a new metric is designed to change the distribution divergence into computable. By minimizing the proposed metric with a min-max training of discrepancy on classifiers and generators, DAT can match noisy data to clean data in the feature space. To the best of our knowledge, DAT is the first to address the noisy label problem from the perspective of the feature distribution. Experiments on synthetic and real-world noisy datasets demonstrate that DAT can consistently outperform other state-of-the-art methods. Codes are available at https://github.com/Tyqnn0323/DAT.

In the biomedical domain, there is an abundance of dense, complex data where objects of interest may be challenging to detect or constrained by limits of human knowledge. Labelled domain specific datasets for supervised tasks are often expensive to obtain, and furthermore discovery of novel distinct objects may be desirable for unbiased scientific discovery. Therefore, we propose leveraging the wealth of annotations in benchmark computer vision datasets to conduct unsupervised instance segmentation for diverse biomedical datasets. The key obstacle is thus overcoming the large domain shift from common to biomedical images. We propose a Domain Adaptive Region-based Convolutional Neural Network (DARCNN), that adapts knowledge of object definition from COCO, a large labelled vision dataset, to multiple biomedical datasets. We introduce a domain separation module, a self-supervised representation consistency loss, and an augmented pseudo-labelling stage within DARCNN to effectively perform domain adaptation across such large domain shifts. We showcase DARCNN's performance for unsupervised instance segmentation on numerous biomedical datasets.

This paper presents a detection-aware pre-training (DAP) approach, which leverages only weakly-labeled classification-style datasets (e.g., ImageNet) for pre-training, but is specifically tailored to benefit object detection tasks. In contrast to the widely used image classification-based pre-training (e.g., on ImageNet), which does not include any location-related training tasks, we transform a classification dataset into a detection dataset through a weakly supervised object localization method based on Class Activation Maps to directly pre-train a detector, making the pre-trained model location-aware and capable of predicting bounding boxes. We show that DAP can outperform the traditional classification pre-training in terms of both sample efficiency and convergence speed in downstream detection tasks including VOC and COCO. In particular, DAP boosts the detection accuracy by a large margin when the number of examples in the downstream task is small.

Semantic segmentation of nighttime images plays an equally important role as that of daytime images in autonomous driving, but the former is much more challenging due to poor illuminations and arduous human annotations. In this paper, we propose a novel domain adaptation network (DANNet) for nighttime semantic segmentation without using labeled nighttime image data. It employs an adversarial training with a labeled daytime dataset and an unlabeled dataset that contains coarsely aligned day-night image pairs. Specifically, for the unlabeled day-night image pairs, we use the pixel-level predictions of static object categories on a daytime image as a pseudo supervision to segment its counterpart nighttime image. We further design a re-weighting strategy to handle the inaccuracy caused by misalignment between day-night image pairs and wrong predictions of daytime images, as well as boost the prediction accuracy of small objects. The proposed DANNet is the first one stage adaptation framework for nighttime semantic segmentation, which does not train additional day-night image transfer models as a separate pre-processing stage. Extensive experiments on Dark Zurich and Nighttime Driving datasets show that our method achieves state-of-the-art performance for nighttime semantic segmentation.

We present the first single-view 3D reconstruction network aimed at recovering geometric details from an input image which encompass both topological shape structures and surface features. Our key idea is to train the network to learn a detail disentangled reconstruction consisting of two functions, one implicit field representing the coarse 3D shape and the other capturing the details. Given an input image, our network, coined D^2IM_Net, encodes it into global and local features which are respectively fed into two decoders. The base decoder uses the global features to reconstruct a coarse implicit field, while the detail decoder reconstructs, from the local features, two displacement maps, defined over the front and back sides of the captured object. The final 3D reconstruction is a fusion between the base shape and the displacement maps, with three losses enforcing the recovery of coarse shape, overall structure, and surface details via a novel Laplacian term.

Neural rendering techniques combining machine learning with geometric reasoning have arisen as one of the most promising approaches for synthesizing novel views of a scene from a sparse set of images. Among these, stands out the Neural radiance fields (NeRF), which trains a deep network to map 5D input coordinates (representing spatial location and viewing direction) into a volume density and view-dependent emitted radiance. However, despite achieving an unprecedented level of photorealism on the generated images, NeRF is only applicable to static scenes, where the same spatial location can be queried from different images. In this paper we introduce D-NeRF, a method that extends neural radiance fields to a dynamic domain, allowing to reconstruct and render novel images of objects under rigid and non-rigid motions. For this purpose we consider time as an additional input to the system, and split the learning process in two main stages: one that encodes the scene into a canonical space and another that maps this canonical representation into the deformed scene at a particular time. Both mappings are learned using fully-connected networks. Once the networks are trained, D-NeRF can render novel images, controlling both the camera view and the time variable, and thus, the object movement. We demonstrate the effectiveness of our approach on scenes with objects under rigid, articulated and non-rigid motions.

State-of-the-art methods for large-scale driving-scene LiDAR segmentation often project the point clouds to 2D space and then process them via 2D convolution. Although this corporation shows the competitiveness in the point cloud, it inevitably alters and abandons the 3D topology and geometric relations. A natural remedy is to utilize the 3D voxelization and 3D convolution network. However, we found that in the outdoor point cloud, the improvement obtained in this way is quite limited. An important reason is the property of the outdoor point cloud, namely sparsity and varying density. Motivated by this investigation, we propose a new framework for the outdoor LiDAR segmentation, where cylindrical partition and asymmetrical 3D convolution networks are designed to explore the 3D geometric pattern while maintaining these inherent properties. Moreover, a point-wise refinement module is introduced to alleviate the interference of lossy voxel-based label encoding. We evaluate the proposed model on two large-scale datasets , i.e., SemanticKITTI and nuScenes. Our method achieves the 1st place in the leaderboard of SemanticKITTI and outperforms existing methods on nuScenes with a noticeable margin. Furthermore, the proposed 3D framework also generalizes well to LiDAR panoptic segmentation and LiDAR 3D detection.

There are rich synchronized audio and visual events in our daily life. Inside the events, audio scenes are associated with the corresponding visual objects; meanwhile, sounding objects can indicate and help to separate their individual sounds in the audio track. Based on this observation, in this paper, we propose a cyclic co-learning (CCoL) paradigm that can jointly learn sounding object visual grounding and audio-visual sound separation in a unified framework. Concretely, we can leverage grounded object-sound relations to improve the results of sound separation. Meanwhile, benefiting from discriminative information from separated sounds, we improve training example sampling for sounding object grounding, which builds a co-learning cycle for the two tasks and makes them mutually beneficial. Extensive experiments show that the proposed framework outperforms the compared recent approaches on both tasks, and they can benefit from each other with our cyclic co-learning. The source code and pre-trained models are released in https://github.com/YapengTian/CCOL-CVPR21.

In this paper, we present a novel unpaired point cloud completion network, named Cycle4Completion, to infer the complete geometries from a partial 3D object. Previous unpaired completion methods merely focus on the learning of geometric correspondence from incomplete shapes to complete shapes, and ignore the learning in the reverse direction, which makes them suffer from low completion accuracy due to the limited 3D shape understanding ability. To address this problem, we propose two simultaneous cycle transformations between the latent spaces of complete shapes and incomplete ones. Specifically, the first cycle transforms shapes from incomplete domain to complete domain, and then projects them back to the incomplete domain. This process learns the geometric characteristic of complete shapes, and maintains the shape consistency between the complete prediction and the incomplete input. Similarly, the inverse cycle transformation starts from complete domain to incomplete domain, and goes back to complete domain to learn the characteristic of incomplete shapes. We experimentally show that our model with the learned bidirectional geometry correspondence outperforms state-of-the-art unpaired completion methods. Code will be available at https://github.com/diviswen/Cycle4Completion.

We aim at constructing a high performance model for defect detection that detects unknown anomalous patterns of an image without anomalous data. To this end, we propose a two-stage framework for building anomaly detectors using normal training data only. We first learn self-supervised deep representations and then build a generative one-class classifier on learned representations. We learn representations by classifying normal data from the CutPaste, a simple data augmentation strategy that cuts an image patch and pastes at a random location of a large image. Our empirical study on MVTec anomaly detection dataset demonstrates the proposed algorithm is general to be able to detect various types of real-world defects. We bring the improvement upon previous arts by 3.1 AUCs when learning representations from scratch. By transfer learning on pretrained representations on ImageNet, we achieve a new state-of-the-art 96.6 AUC. Lastly, we extend the framework to learn and extract representations from patches to allow localizing defective areas without annotations during training.

In this paper we address multi-target domain adaptation (MTDA), where given one labeled source dataset and multiple unlabeled target datasets that differ in data distributions, the task is to learn a robust predictor for all the target domains. We identify two key aspects that can help to alleviate multiple domain-shifts in the MTDA: feature aggregation and curriculum learning. To this end, we propose Curriculum Graph Co-Teaching (CGCT) that uses a dual classifier head, with one of them being a graph convolutional network (GCN) which aggregates features from similar samples across the domains. To prevent the classifiers from over-fitting on its own noisy pseudo-labels we develop a co-teaching strategy with the dual classifier head that is assisted by curriculum learning to obtain more reliable pseudo-labels. Furthermore, when the domain labels are available, we propose Domain-aware Curriculum Learning (DCL), a sequential adaptation strategy that first adapts on the easier target domains, followed by the harder ones. We experimentally demonstrate the effectiveness of our proposed frameworks on several benchmarks and advance the state-of-the-art in the MTDA by large margins (e.g. +5.6% on the DomainNet).

Humans perceive and construct the surrounding world as an arrangement of simple parametric models. In particular, man-made environments commonly consist of volumetric primitives such as cuboids or cylinders. Inferring these primitives is an important step to attain high-level, abstract scene descriptions. Previous approaches directly estimate shape parameters from a 2D or 3D input, and are only able to reproduce simple objects, yet unable to accurately parse more complex 3D scenes. In contrast, we propose a robust estimator for primitive fitting, which can meaningfully abstract real-world environments using cuboids. A RANSAC estimator guided by a neural network fits these primitives to 3D features, such as a depth map. We condition the network on previously detected parts of the scene, thus parsing it one-by-one. To obtain 3D features from a single RGB image, we additionally optimise a feature extraction CNN in an end-to-end manner. However, naively minimising point-to-primitive distances leads to large or spurious cuboids occluding parts of the scene behind. We thus propose an occlusion-aware distance metric correctly handling opaque scenes. The proposed algorithm does not require labour-intensive labels, such as cuboid annotations, for training. Results on the challenging NYU Depth v2 dataset demonstrate that the proposed algorithm successfully abstracts cluttered real-world 3D scene layouts.

Jigsaw puzzle solving, the problem of constructing a coherent whole from a set of non-overlapping unordered fragments, is fundamental to numerous applications, and yet most of the literature has focused thus far on less realistic puzzles whose pieces are identical squares. Here we formalize a new type of jigsaw puzzle where the pieces are general convex polygons generated by cutting through a global polygonal shape with an arbitrary number of straight cuts. We analyze the theoretical properties of such puzzles, including the inherent challenges in solving them once pieces are contaminated with geometrical noise. To cope with such difficulties and obtain tractable solutions, we abstract the problem as a multi-body spring-mass dynamical system endowed with hierarchical loop constraints and a layered reconstruction process that is guided by the pictorial content of the pieces. We define evaluation metrics and present experimental results on both apictorial and pictorial puzzles to indicate that they are solvable completely automatically.

Panoptic segmentation unifies semantic segmentation and instance segmentation which has been attracting increasing attention in recent years. On the other hand, most existing research was conducted under a supervised learning setup whereas domain adaptive panoptic segmentation which is critical in different tasks and applications is largely neglected. We design a domain adaptive panoptic segmentation network that exploits inter-style consistency and inter-task regularization for optimal domain adaptive panoptic segmentation. The inter-style consistency leverages geometric invariance across the same image of the different styles which ` fabricates' certain self-supervisions to guide the network to learn domain-invariant features. The inter-task regularization exploits the complementary nature of instance segmentation and semantic segmentation and uses it as a constraint for better feature alignment across domains. Extensive experiments over multiple domain adaptive panoptic segmentation tasks (e.g. synthetic-to-real and real-to-real) show that our proposed network achieves superior segmentation performance as compared with the state-of-the-art.