CVPR 2022 Papers

Current Image captioning (IC) methods predict textual words sequentially based on the input visual information from the visual feature extractor and the partially generated sentence information. However, for most cases, the partially generated sentence may dominate the target word prediction due to the insufficiency of visual information, making the generated descriptions irrelevant to the content of the given image. In this paper, we propose a Dual Information Flow Network (DIFNet) to address this issue, which takes segmentation feature as another visual information source to enhance the contribution of visual information for prediction. To maximize the use of two information flows, we also propose an effective feature fusion module termed Iterative Independent Layer Normalization (IILN) which can condense the most relevant inputs while retraining modality-specific information in each flow. Experiments show that our method is able to enhance the dependence of prediction on visual information, making word prediction more focused on the visual content, and thus achieve new state-of-the-art performance on the MSCOCO dataset, e.g., 136.2 CIDEr on COCO Karpathy test split.

Monocular 6D pose estimation is a fundamental task in computer vision. Existing works often adopt a twostage pipeline by establishing correspondences and utilizing a RANSAC algorithm to calculate 6 degrees-of-freedom (6DoF) pose. Recent works try to integrate differentiable RANSAC algorithms to achieve an end-to-end 6D pose estimation. However, most of them hardly consider the geometric features in 3D space, and ignore the topology cues when performing differentiable RANSAC algorithms. To this end, we proposed a Depth-Guided Edge Convolutional Network (DGECN) for 6D pose estimation task. We have made efforts from the following three aspects: 1) We take advantages of estimated depth information to guide both the correspondences-extraction process and the cascaded differentiable RANSAC algorithm with geometric information. 2)We leverage the uncertainty of the estimated depth map to improve accuracy and robustness of the output 6D pose. 3) We propose a differentiable Perspective-n-Point(PnP) algorithm via edge convolution to explore the topology relations between 2D-3D correspondences. Experiments demonstrate that our proposed network outperforms current works on both effectiveness and efficiency.

Recent self-supervised pretraining methods for object detection largely focus on pretraining the backbone of the object detector, neglecting key parts of detection architecture. Instead, we introduce DETReg, a new self-supervised method that pretrains the entire object detection network, including the object localization and embedding components. During pretraining, DETReg predicts object localizations to match the localizations from an unsupervised region proposal generator and simultaneously aligns the corresponding feature embeddings with embeddings from a self-supervised image encoder. We implement DETReg using the DETR family of detectors and show that it improves over competitive baselines when finetuned on COCO, PASCAL VOC, and Airbus Ship benchmarks. In low-data regimes, including semi-supervised and few-shot learning settings, DETReg establishes many state-of-the-art results, e.g., on COCO we see a +6.0 AP improvement for 10-shot detection and +3.5 AP improvement when training with only 1% of the labels.

Self- and cross-attention in Transformers provide for high model capacity, making them viable models for object detection. However, Transformers still lag in performance behind CNN-based detectors. This is, we believe, because: (a) Cross-attention is used for both classification and bounding-box regression tasks; (b) Transformer's decoder poorly initializes content queries; and (c) Self-attention poorly accounts for certain prior knowledge which could help improve inductive bias. These limitations are addressed with the corresponding three contributions. First, we propose a new Detection Split Transformer (DESTR) that separates estimation of cross-attention into two independent branches -- one tailored for classification and the other for box regression. Second, we use a mini-detector to initialize the content queries in the decoder with classification and regression embeddings of the respective heads in the mini-detector. Third, we augment self-attention in the decoder to additionally account for pairs of adjacent object queries. Our experiments on the MS-COCO dataset show that DESTR outperforms DETR and its successors.

Backdoor attack is a type of serious security threat to deep learning models.An adversary can provide users with a model trained on poisoned data to manipulate prediction behavior in test stage using a backdoor. The backdoored models behave normally on clean images, yet can be activated and output incorrect prediction if the input is stamped with a specific trigger pattern.Most existing backdoor attacks focus on manually defining imperceptible triggers in input space without considering the abnormality of triggers' latent representations in the poisoned model.These attacks are susceptible to backdoor detection algorithms and even visual inspection.In this paper, We propose a novel and stealthy backdoor attack - DEFEAT. It poisons the clean data using adaptive imperceptible perturbation and restricts latent representation during training process to strengthen our attack's stealthiness and resistance to defense algorithms.We conduct extensive experiments on multiple image classifiers using real-world datasets to demonstrate that our attack can 1) hold against the state-of-the-art defenses, 2) deceive the victim model with high attack success without jeopardizing model utility, and 3) provide practical stealthiness on image data.

Deep learning has become an increasingly popular and powerful methodology for modern pattern recognition systems. However, many deep neural networks have millions or billions of parameters, making them untenable for real-world applications due to constraints on memory size or latency requirements. As a result, efficient network compression techniques are often required for a widespread adoption of deep learning methods. We present DECORE, a reinforcement learning based approach to automate the network compression process. DECORE assigns an agent to each channel in the network along with a light policy gradient method to learn which neurons or channels to be kept or removed. Each agent in the network has just one parameter (keep or drop) to learn, which leads to a much faster training process compared to existing approaches. DECORE also gives state-of-the-art compression results on various network architectures and various datasets. For example, on the ResNet-110 architecture, DECORE achieves a 64.8% compression rate and 61.8% FLOPs reduction as compared to the baseline model without any major accuracy loss on the CIFAR-10 dataset. It can reduce the size of regular architectures like the VGG network by up to 99% with just a small accuracy drop of 2.28%. For a larger dataset like ImageNet it can compress the ResNet-50 architecture by 44.7% and reduces FLOPs by 42.3%, with just a 0.69% drop on Top-5 accuracy of the uncompressed model. We also demonstrate that DECORE can be used to search for compressed network architectures based on various constraints, such as memory and FLOPs.

Consistency-based Semi-supervised learning (SSL) has achieved promising performance recently. However, the success largely depends on the assumption that the labeled and unlabeled data share an identical class distribution, which is hard to meet in real practice. The distribution mismatch between the labeled and unlabeled sets can cause severe bias in the pseudo-labels of SSL, resulting in significant performance degradation. To bridge this gap, we put forward a new SSL learning framework, named Distribution Consistency SSL (DC-SSL), which rectifies the pseudo-labels from a distribution perspective. The basic idea is to directly estimate a reference class distribution (RCD), which is regarded as a surrogate of the ground truth class distribution about the unlabeled data, and then improve the pseudo-labels by encouraging the predicted class distribution (PCD) of the unlabeled data to approach RCD gradually. To this end, this paper revisits the Exponentially Moving Average (EMA) model and utilizes it to estimate RCD in an iteratively improved manner, which is achieved with a momentum-update scheme throughout the training procedure. On top of this, two strategies are proposed for RCD to rectify the pseudo-label prediction, respectively. They correspond to an efficient training-free scheme and a training-based alternative that generates more accurate and reliable predictions. DC-SSL is evaluated on multiple SSL benchmarks and demonstrates remarkable performance improvement over competitive methods under matched- and mismatched-distribution scenarios.

Automatic tooth instance segmentation on 3D dental models is a fundamental task for computer-aided orthodontic treatments. Existing learning-based methods rely heavily on expensive point-wise annotations. To alleviate this problem, we are the first to explore a low-cost annotation way for 3D tooth instance segmentation, i.e., labeling all tooth centroids and only a few teeth for each dental model. Regarding the challenge when only weak annotation is provided, we present a dental arch prior-assisted 3D tooth segmentation method, namely DArch. Our DArch consists of two stages, including tooth centroid detection and tooth instance segmentation. Accurately detecting the tooth centroids can help locate the individual tooth, thus benefiting the segmentation. Thus, our DArch proposes to leverage the dental arch prior to assist the detection. Specifically, we firstly propose a coarse-to-fine method to estimate the dental arch, in which the dental arch is initially generated by Bezier curve regression and then a lightweight network is trained to refine it. With the estimated dental arch, we then propose a novel Arch-aware Point Sampling (APS) method to assist the tooth centroid proposal generation. Meantime, a segmentor is independently trained using a patch-based training strategy, aiming to segment a tooth instance from a 3D patch centered at the tooth centroid. Experimental results on 4,773 dental models have shown our DArch can accurately segment each tooth of a dental model, and its performance is superior to the state-of-the-art methods.

The paradigm of training models on massive data without label through self-supervised learning (SSL) and finetuning on many downstream tasks has become a trend recently. However, due to the high training costs and the unconsciousness of downstream usages, most self-supervised learning methods lack the capability to correspond to the diversities of downstream scenarios, as there are various data domains, latency constraints and etc. Neural architecture search (NAS) is one universally acknowledged fashion to conquer the issues above, but applying NAS on SSL seems impossible as there is no label or metric provided for judging model selection. In this paper, we present DATA, a simple yet effective NAS approach specialized for SSL that provides Domain-Aware and Task-Aware pre-training. Specifically, we (i) train a supernet which could be deemed as a set of millions of networks covering a wide range of model scales without any label, (ii) propose a flexible searching mechanism compatible with SSL that enables finding networks of different computation costs, for various downstream vision tasks and data domains without explicit metric provided. Instantiated With MoCov2, our method achieves promising results across a wide range of computation costs on downstream tasks, including image classification, object detection and semantic segmentation. DATA is orthogonal to most existing SSL methods and endows them the ability of customization on downstream needs. Extensive experiments on other SSL methods, including BYOL, ReSSL and DenseCL demonstrate the generalizability of the proposed method. Code would be made available soon.

The capability of the traditional semi-supervised learning (SSL) methods is far from real-world application due to severely biased pseudo-labels caused by (1) class imbalance and (2) class distribution mismatch between labeled and unlabeled data. This paper addresses such a relatively under-explored problem. First, we propose a general pseudo-labeling framework that class-adaptively blends the semantic pseudo-label from a similarity-based classifier to the linear one from the linear classifier, after making the observation that both types of pseudo-labels have complementary properties in terms of bias. We further introduce a novel semantic alignment loss to establish balanced feature representation to reduce the biased predictions from the classifier. We term the whole framework as Distribution-Aware Semantics-Oriented (DASO) Pseudo-label. We conduct extensive experiments in a wide range of imbalanced benchmarks: CIFAR10/100-LT, STL10-LT, and large-scale long-tailed Semi-Aves with open-set class, and demonstrate that, the proposed DASO framework reliably improves SSL learners with unlabeled data especially when both (1) class imbalance and (2) distribution mismatch dominate.

Autonomous driving faces great safety challenges for a lack of global perspective and the limitation of long-range perception capabilities. It has been widely agreed that vehicle-infrastructure cooperation is required to achieve Level 5 autonomy. However, there is still NO dataset from real scenarios available for computer vision researchers to work on vehicle-infrastructure cooperation-related problems. To accelerate computer vision research and innovation for Vehicle-Infrastructure Cooperative Autonomous Driving (VICAD), we release DAIR-V2X Dataset, which is the first large-scale, multi-modal, multi-view dataset from real scenarios for VICAD. DAIR-V2X comprises 71254 LiDAR frames and 71254 Camera frames, and all frames are captured from real scenes with 3D annotations. The Vehicle-Infrastructure Cooperative 3D Object Detection problem (VIC3D) is introduced, formulating the problem of collaboratively locating and identifying 3D objects using sensory input from both vehicles and infrastructure. In addition to solving traditional 3D object detection problems, the solution of VIC3D needs to consider the time asynchrony problem between vehicle and infrastructure sensors and the data transmission cost between them. Furthermore, we propose Time Compensation Late Fusion (TCLF), a late fusion framework for the VIC3D task as a benchmark based on DAIR-V2X. Find data, code, and more up-to-date information at \href https://thudair.baai.ac.cn/index https://thudair.baai.ac.cn/index and \href https://github.com/AIR-THU/DAIR-V2X https://github.com/AIR-THU/DAIR-V2X .

As acquiring pixel-wise annotations of real-world images for semantic segmentation is a costly process, a model can instead be trained with more accessible synthetic data and adapted to real images without requiring their annotations. This process is studied in unsupervised domain adaptation (UDA). Even though a large number of methods propose new adaptation strategies, they are mostly based on outdated network architectures. As the influence of recent network architectures has not been systematically studied, we first benchmark different network architectures for UDA and newly reveal the potential of Transformers for UDA semantic segmentation. Based on the findings, we propose a novel UDA method, DAFormer. The network architecture of DAFormer consists of a Transformer encoder and a multi-level context-aware feature fusion decoder. It is enabled by three simple but crucial training strategies to stabilize the training and to avoid overfitting to the source domain: While (1) Rare Class Sampling on the source domain improves the quality of the pseudo-labels by mitigating the confirmation bias of self-training toward common classes, (2) a Thing-Class ImageNet Feature Distance and (3) a learning rate warmup promote feature transfer from ImageNet pretraining. DAFormer represents a major advance in UDA. It improves the state of the art by 10.8 mIoU for GTA-to-Cityscapes and 5.4 mIoU for Synthia-to-Cityscapes and enables learning even difficult classes such as train, bus, and truck well. The implementation is available at https://github.com/lhoyer/DAFormer.

We present DAD-3DHeads, a dense and diverse large-scale dataset, and a robust model for 3D Dense Head Alignment in-the-wild. It contains annotations of over 3.5K landmarks that accurately represent 3D head shape compared to the ground-truth scans. The data-driven model, DAD-3DNet, trained on our dataset, learns shape, expression, and pose parameters, and performs 3D reconstruction of a FLAME mesh. The model also incorporates a landmark prediction branch to take advantage of rich supervision and co-training of multiple related tasks. Experimentally, DAD- 3DNet outperforms or is comparable to the state-of-the-art models in (i) 3D Head Pose Estimation on AFLW2000-3D and BIWI, (ii) 3D Face Shape Reconstruction on NoW and Feng, and (iii) 3D Dense Head Alignment and 3D Landmarks Estimation on DAD-3DHeads dataset. Finally, diversity of DAD-3DHeads in camera angles, facial expressions, and occlusions enables a benchmark to study in-the-wild generalization and robustness to distribution shifts. The dataset webpage is https://p.farm/research/dad-3dheads.

We introduce the dynamic grasp synthesis task: given an object with a known 6D pose and a grasp reference, our goal is to generate motions that move the object to a target 6D pose. This is challenging, because it requires reasoning about the complex articulation of the human hand and the intricate physical interaction with the object. We propose a novel method that frames this problem in the reinforcement learning framework and leverages a physics simulation, both to learn and to evaluate such dynamic interactions. A hierarchical approach decomposes the task into low-level grasping and high-level motion synthesis. It can be used to generate novel hand sequences that approach, grasp, and move an object to a desired location, while retaining human-likeness. We show that our approach leads to stable grasps and generates a wide range of motions. Furthermore, even imperfect labels can be corrected by our method to generate dynamic interaction sequences.

Curating a large set of fully annotated training data can be costly, especially for the tasks of medical image segmentation. Scribble, a weaker form of annotation, is more obtainable in practice, but training segmentation models from limited supervision of scribbles is still challenging. To address the difficulties, we propose a new framework for scribble learning-based medical image segmentation, which is composed of mix augmentation and cycle consistency and thus is referred to as CycleMix. For augmentation of supervision, CycleMix adopts the mixup strategy with a dedicated design of random occlusion, to perform increments and decrements of scribbles. For regularization of supervision, CycleMix intensifies the training objective with consistency losses to penalize inconsistent segmentation, which results in significant improvement of segmentation performance. Results on two open datasets, i.e., ACDC and MSCMRseg, showed that the proposed method achieved exhilarating performance, demonstrating comparable or even better accuracy than the fully-supervised methods. The code and expert-made scribble annotations for MSCMRseg are publicly available at https://github.com/BWGZK/CycleMIx.

CounterFactual (CF) visual explanations try to find images similar to the query image that change the decision of a vision system to a specified outcome. Existing methods either require inference-time optimization or joint training with a generative adversarial model which makes them time-consuming and difficult to use in practice. We propose a novel approach, Cycle-Consistent Counterfactuals by Latent Transformations (C3LT), which learns a latent transformation that automatically generates visual CFs by steering in the latent space of generative models. Our method uses cycle consistency between the query and CF latent representations which helps our training to find better solutions. C3LT can be easily plugged into any state-of-the-art pretrained generative network. This enables our method to generate high-quality and interpretable CF images at high resolution such as those in ImageNet. In addition to several established metrics for evaluating CF explanations, we introduce a novel metric tailored to assess the quality of the generated CF examples and validate the effectiveness of our method on an extensive set of experiments.

This paper investigates crowd counting in the frequency domain, which is a novel direction compared to the traditional view in the spatial domain. By transforming the density map into the frequency domain and using the nice properties of the characteristic function, we propose a novel method that is simple, effective, and efficient. The solid theoretical analysis ends up as an implementation-friendly loss function, which requires only standard tensor operations in the training process. We prove that our loss function is an upper bound of the pseudo sup norm metric between the ground truth and the prediction density map (over all of their sub-regions), and demonstrate its efficacy and efficiency versus other loss functions. The experimental results also show its competitiveness to the state-of-the-art on five benchmark data sets: ShanghaiTech A and B, UCF-QNRF, JHU++, and NWPU.

Manual annotation of large-scale point cloud dataset for varying tasks such as 3D object classification, segmentation and detection is often laborious owing to the irregular structure of point clouds. Self-supervised learning, which operates without any human labeling, is a promising approach to address this issue. We observe in the real world that humans are capable of mapping the visual concepts learnt from 2D images to understand the 3D world. Encouraged by this insight, we propose CrossPoint, a simple cross-modal contrastive learning approach to learn transferable 3D point cloud representations. It enables a 3D-2D correspondence of objects by maximizing agreement between point clouds and the corresponding rendered 2D image in the invariant space, while encouraging invariance to transformations in the point cloud modality. Our joint training objective combines the feature correspondences within and across modalities, thus ensembles a rich learning signal from both 3D point cloud and 2D image modalities in a self-supervised fashion. Experimental results show that our approach outperforms the previous unsupervised learning methods on a diverse range of downstream tasks including 3D object classification and segmentation. Further, the ablation studies validates the potency of our approach for a better point cloud understanding. Code and pretrained models are available at https://github.com/MohamedAfham/CrossPoint.

We present a visual localization system that learns to estimate camera poses in the real world with the help of synthetic data. Despite significant progress in recent years, most learning-based approaches to visual localization target at a single domain and require a dense database of geo-tagged images to function well. To mitigate the data scarcity issue and improve the scalability of the neural localization models, we introduce TOPO-DataGen, a versatile synthetic data generation tool that traverses smoothly between the real and virtual world, hinged on the geographic camera viewpoint. New large-scale sim-to-real benchmark datasets are proposed to showcase and evaluate the utility of the said synthetic data. Our experiments reveal that synthetic data generically enhances the neural network performance on real data. Furthermore, we introduce CrossLoc, a cross-modal visual representation learning approach to pose estimation that makes full use of the scene coordinate ground truth via self-supervision. Without any extra data, CrossLoc significantly outperforms the state-of-the-art methods and achieves substantially higher real-data sample efficiency. Our code and datasets are all available at crossloc.github.io.