CVPR 2021 Papers

Prior research on self-supervised learning has led to considerable progress on image classification, but often with degraded transfer performance on object detection. The objective of this paper is to advance self-supervised pretrained models specifically for object detection. Based on the inherent difference between classification and detection, we propose a new self-supervised pretext task, called instance localization. Image instances are pasted at various locations and scales onto background images. The pretext task is to predict the instance category given the composited images as well as the foreground bounding boxes. We show that integration of bounding boxes into pretraining promotes better alignment between convolutional features and region boxes. In addition, we propose an augmentation method on the bounding boxes to further enhance this feature alignment. As a result, our model becomes weaker at Imagenet semantic classification but stronger at image patch localization, with an overall stronger pretrained model for object detection. Experimental results demonstrate that our approach yields state-of-the-art transfer learning results for object detection on PASCAL VOC and MSCOCO.

Domain adaptation deals with training models using large scale labeled data from a specific source domain and then adapting the knowledge to certain target domains that have few or no labels. Many prior works learn domain agnostic feature representations for this purpose using a global distribution alignment objective which does not take into account the finer class specific structure in the source and target domains. We address this issue in our work and propose an instance affinity based criterion for source to target transfer during adaptation, called ILA-DA. We first propose a reliable and efficient method to extract similar and dissimilar samples across source and target, and utilize a multi-sample contrastive loss to drive the domain alignment process. ILA-DA simultaneously accounts for intra-class clustering as well as inter-class separation among the categories, resulting in less noisy classifier boundaries, improved transferability and increased accuracy. We verify the effectiveness of ILA-DA by observing consistent improvements in accuracy over popular domain adaptation approaches on a variety of benchmark datasets and provide insights into the proposed alignment approach. Code will be made publicly available at https://github.com/astuti/ILA-DA.

Cross-domain weakly supervised object detection aims to adapt object-level knowledge from a fully labeled source domain dataset (i.e. with object bounding boxes) to train object detectors for target domains that are weakly labeled (i.e. with image-level tags). Instead of domain-level distribution matching, as popularly adopted in the literature, we propose to learn pixel-wise cross-domain correspondences for more precise knowledge transfer. It is realized through a novel cross-domain co-attention scheme trained as region competition. In this scheme, the cross-domain correspondence module seeks for informative features on the target domain image, which after being warped to the source domain image, could best explain its annotations. Meanwhile, a collaborative mask generator competes to mask out the relevant target image region to make the remaining features uninformative. Such competitive learning strives to correlate the full foreground in cross-domain image pairs, revealing the accurate object extent in target domain. To alleviate the ambiguity of inter-domain correspondence learning, a domain-cycle consistency regularizer is futher proposed to leverage the more reliable intra-domain correspondence. The proposed approach achieves consistent improvements over existing approaches by a considerable margin, demonstrated by the experiments on various datasets.

We study image segmentation from an information-theoretic perspective, proposing a novel adversarial method that performs unsupervised segmentation by partitioning images into maximally independent sets. More specifically, we group image pixels into foreground and background, with the goal of minimizing predictability of one set from the other. An easily computed loss drives a greedy search process to maximize inpainting error over these partitions. Our method does not involve training deep networks, is computationally cheap, class-agnostic, and even applicable in isolation to a single unlabeled image. Experiments demonstrate that it achieves a new state-of-the-art in unsupervised segmentation quality, while being substantially faster and more general than competing approaches.

Improving the performance of face forgery detectors often requires more identity-swapped images of higher-quality. One core objective of identity swapping is to generate identity-discriminative faces that are distinct from the target while identical to the source. To this end, properly disentangling identity and identity-irrelevant information is critical and remains a challenging endeavor. In this work, we propose a novel information disentangling and swapping network, called InfoSwap, to extract the most expressive information for identity representation from a pre-trained face recognition model. The key insight of our method is to formulate the learning of disentangled representations as optimizing an information bottleneck trade-off, in terms of finding an optimal compression of the pre-trained latent features. Moreover, a novel identity contrastive loss is proposed for further disentanglement by requiring a proper distance between the generated identity and the target. While the most prior works have focused on using various loss functions to implicitly guide the learning of representations, we demonstrate that our model can provide explicit supervision for learning disentangled representations, achieving impressive performance in generating more identity-discriminative swapped faces.

In computer-aided design (CAD), the ability to "reverse engineer" the modeling steps used to create 3D shapes is a long-sought-after goal. This process can be decomposed into two sub-problems: converting an input mesh or point cloud into a boundary representation (or B-rep), and then inferring modeling operations which construct this B-rep. In this paper, we present a new system for solving the second sub-problem. Central to our approach is a new geometric representation: the zone graph. Zones are the set of solid regions formed by extending all B-Rep faces and partitioning space with them; a zone graph has these zones as its nodes, with edges denoting geometric adjacencies between them. Zone graphs allow us to tractably work with industry-standard CAD operations, unlike prior work using CSG with parametric primitives. We focus on CAD programs consisting of sketch + extrude + Boolean operations, which are common in CAD practice. We phrase our problem as search in the space of such extrusions permitted by the zone graph, and we train a graph neural network to score potential extrusions in order to accelerate the search. We show that our approach outperforms an existing CSG inference baseline in terms of geometric reconstruction accuracy and reconstruction time, while also creating more plausible modeling sequences.

Indoor panorama typically consists of human-made structures parallel or perpendicular to gravity. We leverage this phenomenon to approximate the scene in a 360-degree image with (H)orizontal-planes and (V)ertical-planes. To this end, we propose an effective divide-and-conquer strategy that divides pixels based on their plane orientation estimation; then, the succeeding instance segmentation module conquers the task of planes clustering more easily in each plane orientation group. Besides, parameters of V-planes depend on camera yaw rotation, but translation-invariant CNNs are less aware of the yaw change. We thus propose a yaw-invariant V-planar reparameterization for CNNs to learn. We create a benchmark for indoor panorama planar reconstruction by extending existing 360 depth datasets with ground truth H&V-planes (referred to as "PanoH&V" dataset) and adopt state-of-the-art planar reconstruction methods to predict H&V-planes as our baselines. Our method outperforms the baselines by a large margin on the proposed dataset.

Image-based methods for indoor lighting estimation suffer from the problem of intensity-distance ambiguity. This paper introduces a novel setup to help alleviate the ambiguity based on the event camera. We further demonstrate that estimating the distance of a light source becomes a well-posed problem under this setup, based on which an optimization-based method and a learning-based method are proposed. Our experimental results validate that our approaches not only achieve superior performance for indoor lighting estimation (especially for the close light) but also significantly alleviate the intensity-distance ambiguity.

Current deep learning architectures suffer from catastrophic forgetting, a failure to retain knowledge of previously learned classes when incrementally trained on new classes. The fundamental roadblock faced by deep learning methods is that the models are optimized as "black boxes", making it difficult to properly adjust the model parameters to preserve knowledge about previously seen data. To overcome the problem of catastrophic forgetting, we propose utilizing an alternative "white box" architecture derived from the principle of rate reduction, where each layer of the network is explicitly computed without back propagation. Under this paradigm, we demonstrate that, given a pretrained network and new data classes, our approach can provably construct a new network that emulates joint training with all past and new classes. Finally, our experiments show that our proposed learning algorithm observes significantly less decay in classification performance, outperforming state of the art methods on MNIST and CIFAR-10 by a large margin and justifying the use of "white box" algorithms for incremental learning even for sufficiently complex image data.

Few-shot instance segmentation methods are promising when labeled training data for novel classes is scarce. However, current approaches do not facilitate flexible addition of novel classes. They also require that examples of each class are provided at train and test time, which is memory intensive. In this paper, we address these limitations by presenting the first incremental approach to few-shot instance segmentation: iMTFA. We learn discriminative embeddings for object instances that are merged into class representatives. Storing embedding vectors rather than images effectively solves the memory overhead problem. We match these class embeddings at the RoI-level using cosine similarity. This allows us to add new classes without the need for further training or access to previous training data. In a series of experiments, we consistently outperform the current state-of-the-art. Moreover, the reduced memory requirements allow us to evaluate, for the first time, few-shot instance segmentation performance on all classes in COCO jointly.

As a variant of standard convolution, a dilated convolution can control effective receptive fields and handle large scale variance of objects without introducing additional computational costs. To fully explore the potential of dilated convolution, we proposed a new type of dilated convolution (referred to as inception convolution), where the convolution operations have independent dilation patterns among different axes, channels and layers. To develop a practical method for learning complex inception convolution based on the data, a simple but effective search algorithm, referred to as efficient dilation optimization (EDO), is developed. Based on statistical optimization, the EDO method operates in a low-cost manner and is extremely fast when it is applied on large scale datasets. Empirical results validate that our method achieves consistent performance gains for image recognition, object detection, instance segmentation, human detection, and human pose estimation. For instance, by simply replacing the 3 x 3 standard convolution in the ResNet-50 backbone with inception convolution, we significantly improve the AP of Faster R-CNN from 36.4% to 39.2% on MS COCO.

Style transfer aims to render the content of a given image in the graphical/artistic style of another image. The fundamental concept underlying Neural Style Transfer (NST) is to interpret style as a distribution in the feature space of a Convolutional Neural Network, such that a desired style can be achieved by matching its feature distribution. We show that most current implementations of that concept have important theoretical and practical limitations, as they only partially align the feature distributions. We propose a novel approach that matches the distributions more precisely, thus reproducing the desired style more faithfully, while still being computationally efficient. Specifically, we adapt the dual form of Central Moment Discrepancy, as recently proposed for domain adaptation, to minimize the difference between the target style and the feature distribution of the output image. The dual interpretation of this metric explicitly matches all higher-order centralized moments and is therefore a natural extension of existing NST methods that only take into account the first and second moments. Our experiments confirm that the strong theoretical properties also translate to visually better style transfer, and better disentangle style from semantic image content.

Although deep neural networks (DNNs) have achieved tremendous performance in diverse vision challenges, they are surprisingly susceptible to adversarial examples, which are born of intentionally perturbing benign samples in a human-imperceptible fashion. It thus poses security concerns on the deployment of DNNs in practice, particularly in safety- and security-sensitive domains. To investigate the robustness of DNNs, transfer-based attacks have attracted a growing interest recently due to their high practical applicability, where attackers craft adversarial samples with local models and employ the resultant samples to attack a remote black-box model. However, existing transfer-based attacks frequently suffer from low success rates due to overfitting to the adopted local model. To boost the transferability of adversarial samples, we propose to improve the robustness of synthesized adversarial samples via adversarial transformations. Specifically, we employ an adversarial transformation network to model the most harmful distortions that can destroy adversarial noises and require the synthesized adversarial samples to become resistant to such adversarial transformations. Extensive experiments on the ImageNet benchmark showcase the superiority of our method to state-of-the-art baselines in attacking both undefended and defended models.

Deepfakes is a branch of malicious techniques that transplant a target face to the original one in videos, resulting in serious problems such as infringement of copyright, confusion of information, or even public panic. Previous efforts for Deepfakes videos detection mainly focused on appearance features, which have a risk of being bypassed by sophisticated manipulation, also resulting in high model complexity and sensitiveness to noise. Besides, how to mine the temporal features of manipulated videos and exploit them is still an open question. We propose an efficient and robust framework named LRNet for detecting Deepfakes videos through temporal modeling on precise geometric features. A novel calibration module is devised to enhance the precision of geometric features, making it more discriminative, and a two-stream Recurrent Neural Network (RNN) is constructed for sufficient exploitation of temporal features. Compared to previous methods, our proposed method is lighter-weighted and easier to train. Moreover, our method has shown robustness in detecting highly compressed or noise corrupted videos. Our model achieved 0.999 AUC on FaceForensics++ dataset. Meanwhile, it has a graceful decline in performance (-0.042 AUC) when faced with highly compressed videos.

Weakly supervised phrase grounding aims at learning region-phrase correspondences using only image-sentence pairs. A major challenge thus lies in the missing links between image regions and sentence phrases during training. To address this challenge, we leverage a generic object detector at training time, and propose a contrastive learning framework that accounts for both region-phrase and image-sentence matching. Our core innovation is the learning of a region-phrase score function, based on which an image-sentence score function is further constructed. Importantly, our region-phrase score function is learned by distilling from soft matching scores between the detected object names and candidate phrases within an image-sentence pair, while the image-sentence score function is supervised by ground-truth image-sentence pairs. The design of such score functions removes the need of object detection at test time, thereby significantly reducing the inference cost. Without bells and whistles, our approach achieves state-of-the-art results on visual phrase grounding, surpassing previous methods that require expensive object detectors at test time.

Unsupervised image clustering methods often introduce alternative objectives to indirectly train the model and are subject to faulty predictions and overconfident results. To overcome these challenges, the current research proposes an innovative model RUC that is inspired by robust learning. RUC's novelty is at utilizing pseudo-labels of existing image clustering models as a noisy dataset that may include misclassified samples. Its retraining process can revise misaligned knowledge and alleviate the overconfidence problem in predictions. The model's flexible structure makes it possible to be used as an add-on module to other clustering methods and helps them achieve better performance on multiple datasets. Extensive experiments show that the proposed model can adjust the model confidence with better calibration and gain additional robustness against adversarial noise.

Face recognition is greatly improved by deep convolutional neural networks (CNNs). Recently, these face recognition models have been used for identity authentication in security sensitive applications. However, deep CNNs are vulnerable to adversarial patches, which are physically realizable and stealthy, raising new security concerns on the real-world applications of these models. In this paper, we evaluate the robustness of face recognition models using adversarial patches based on transferability, where the attacker has limited accessibility to the target models. First, we extend the existing transfer-based attack techniques to generate transferable adversarial patches. However, we observe that the transferability is sensitive to initialization and degrades when the perturbation magnitude is large, indicating the overfitting to the substitute models. Second, we propose to regularize the adversarial patches on the low dimensional data manifold. The manifold is represented by generative models pre-trained on legitimate human face images. Using face-like features as adversarial perturbations through optimization on the manifold, we show that the gaps between the responses of substitute models and the target models dramatically decrease, exhibiting a better transferability. Extensive digital world experiments are conducted to demonstrate the superiority of the proposed method in the black-box setting. We apply the proposed method in the physical world as well.

Despite existing pioneering works on sign language translation (SLT), there is a non-trivial obstacle, i.e., the limited quantity of parallel sign-text data. To tackle this parallel data bottleneck, we propose a sign back-translation (SignBT) approach, which incorporates massive spoken language texts into SLT training. With a text-to-gloss translation model, we first back-translate the monolingual text to its gloss sequence. Then, the paired sign sequence is generated by splicing pieces from an estimated gloss-to-sign bank at the feature level. Finally, the synthetic parallel data serves as a strong supplement for the end-to-end training of the encoder-decoder SLT framework. To promote the SLT research, we further contribute CSL-Daily, a large-scale continuous SLT dataset. It provides both spoken language translations and gloss-level annotations. The topic revolves around people's daily lives (e.g., travel, shopping, medical care), the most likely SLT application scenario. Extensive experimental results and analysis of SLT methods are reported on CSL-Daily. With the proposed sign back-translation method, we obtain a substantial improvement over previous state-of-the-art SLT methods.

Crop-based training strategies decouple training resolution from GPU memory consumption, allowing the use of large-capacity panoptic segmentation networks on multi-megapixel images. Using crops, however, can introduce a bias towards truncating or missing large objects. To address this, we propose a novel crop-aware bounding box regression loss (CABB loss), which promotes predictions to be consistent with the visible parts of the cropped objects, while not over-penalizing them for extending outside of the crop. We further introduce a novel data sampling and augmentation strategy which improves generalization across scales by counteracting the imbalanced distribution of object sizes. Combining these two contributions with a carefully designed, top-down panoptic segmentation architecture, we obtain new state-of-the-art results on the challenging Mapillary Vistas (MVD), Indian Driving and Cityscapes datasets, surpassing the previously best approach on MVD by +4.5% PQ and +5.2% mAP.