CVPR 2022 Papers

Existing text-to-image synthesis methods generally are only applicable to words in the training dataset. However, human faces are so variable to be described with limited words. So this paper proposes the first free-style text-to-face method namely AnyFace enabling much wider open world applications such as metaverse, social media, cosmetics, forensics, etc. AnyFace has a novel two-stream framework for face image synthesis and manipulation given arbitrary descriptions of the human face. Specifically, one stream performs text-to-face generation and the other conducts face image reconstruction. Facial text and image features are extracted using the CLIP (Contrastive Language-Image Pre-training) encoders. And a collaborative Cross Modal Distillation (CMD) module is designed to align the linguistic and visual features across these two streams. Furthermore, a Diverse Triplet Loss (DT loss) is developed to model fine-grained features and improve facial diversity. Extensive experiments on Multi-modal CelebA-HQ and CelebAText-HQ demonstrate significant advantages of AnyFace over state-of-the-art methods. AnyFace can achieve high-quality, high-resolution, and high-diversity face synthesis and manipulation results without any constraints on the number and content of input captions.

Understanding animals' behaviors is significant for a wide range of applications. However, existing animal behavior datasets have limitations in multiple aspects, including limited numbers of animal classes, data samples and provided tasks, and also limited variations in environmental conditions and viewpoints. To address these limitations, we create a large and diverse dataset, Animal Kingdom, that provides multiple annotated tasks to enable a more thorough understanding of natural animal behaviors. The wild animal footages used in our dataset record different times of the day in extensive range of environments containing variations in backgrounds, viewpoints, illumination and weather conditions. More specifically, our dataset contains 50 hours of annotated videos to localize relevant animal behavior segments in long videos for the video grounding task, 30K video sequences for the fine-grained multi-label action recognition task, and 33K frames for the pose estimation task, which correspond to a diverse range of animals with 850 species across 6 major animal classes. Such a challenging and comprehensive dataset shall be able to facilitate the community to develop, adapt, and evaluate various types of advanced methods for animal behavior analysis. Moreover, we propose a Collaborative Action Recognition (CARe) model that learns general and specific features for action recognition with unseen new animals. This method achieves promising performance in our experiments. Our dataset can be found at https://sutdcv.github.io/Animal-Kingdom.

In the field of computer vision, recent works show that a pure MLP architecture mainly stacked by fully-connected layers can achieve competing performance with CNN and transformer. An input image of vision MLP is usually split into multiple tokens (patches), while the existing MLP models directly aggregate them with fixed weights, neglecting the varying semantic information of tokens from different images. To dynamically aggregate tokens, we propose to represent each token as a wave function with two parts, amplitude and phase. Amplitude is the original feature and the phase term is a complex value changing according to the semantic contents of input images. Introducing the phase term can dynamically modulate the relationship between tokens and fixed weights in MLP. Based on the wave-like token representation, we establish a novel Wave-MLP architecture for vision tasks. Extensive experiments demonstrate that the proposed Wave-MLP is superior to the state-of-the-art MLP architectures on various vision tasks such as image classification, object detection and semantic segmentation. The source code is available at https://github.com/huawei-noah/CV-Backbones/tree/master/wavemlp_pytorch and https://gitee.com/mindspore/models/tree/master/research/cv/wave_mlp.

Nowadays, cameras equipped with AI systems can capture and analyze images to detect people automatically. However, the AI system can make mistakes when receiving deliberately designed patterns in the real world, i.e., physical adversarial examples. Prior works have shown that it is possible to print adversarial patches on clothes to evade DNN-based person detectors. However, these adversarial examples could have catastrophic drops in the attack success rate when the viewing angle (i.e., the camera's angle towards the object) changes. To perform a multi-angle attack, we propose Adversarial Texture (AdvTexture). AdvTexture can cover clothes with arbitrary shapes so that people wearing such clothes can hide from person detectors from different viewing angles. We propose a generative method, named Toroidal-Cropping-based Expandable Generative Attack (TC-EGA), to craft AdvTexture with repetitive structures. We printed several pieces of cloth with AdvTexure and then made T-shirts, skirts, and dresses in the physical world. Experiments showed that these clothes could fool person detectors in the physical world.

The Skinned Multi-Person Linear (SMPL) model represents human bodies by mapping pose and shape parameters to body meshes. However, not all pose and shape parameter values yield physically-plausible or even realistic body meshes. In other words, SMPL is under-constrained and may yield invalid results. We propose learning a prior that restricts the SMPL parameters to values that produce realistic poses via adversarial training. We show that our learned prior covers the diversity of the real-data distribution, facilitates optimization for 3D reconstruction from 2D keypoints, and yields better pose estimates when used for regression from images. For all these tasks, it outperforms the state-of-the-art VAE-based approach to constraining the SMPL parameters. The code will be made available at https://github.com/cvlab-epfl/adv_param_pose_prior.

Single-photon avalanche diodes (SPADs) are growing in popularity for depth sensing tasks. However, SPADs still struggle in the presence of high ambient light due to the effects of pile-up. Conventional techniques leverage fixed or asynchronous gating to minimize pile-up effects, but these gating schemes are all non-adaptive, as they are unable to incorporate factors such as scene priors and previous photon detections into their gating strategy. We propose an adaptive gating scheme built upon Thompson sampling. Adaptive gating periodically updates the gate position based on prior photon observations in order to minimize depth errors. Our experiments show that our gating strategy results in significantly reduced depth reconstruction error and acquisition time, even when operating outdoors under strong sunlight conditions.

Deep learning in the presence of noisy annotations has been studied extensively in classification, but much less in segmentation tasks. In this work, we study the learning dynamics of deep segmentation networks trained on inaccurately-annotated data. We discover a phenomenon that has been previously reported in the context of classification: the networks tend to first fit the clean pixel-level labels during an "early-learning" phase, before eventually memorizing the false annotations. However, in contrast to classification, memorization in segmentation does not arise simultaneously for all semantic categories. Inspired by these findings, we propose a new method for segmentation from noisy annotations with two key elements. First, we detect the beginning of the memorization phase separately for each category during training. This allows us to adaptively correct the noisy annotations in order to exploit early learning. Second, we incorporate a regularization term that enforces consistency across scales to boost robustness against annotation noise. Our method outperforms standard approaches on a medical-imaging segmentation task where noises are synthesized to mimic human annotation errors. It also provides robustness to realistic noisy annotations present in weakly-supervised semantic segmentation, achieving state-of-the-art results on PASCAL VOC 2012. Code is available at https://github.com/Kangningthu/ADELE

Traditional object detectors employ the dense paradigm of scanning over locations and scales in an image. The recent query-based object detectors break this convention by decoding image features with a set of learnable queries. However, this paradigm still suffers from slow convergence, limited performance, and design complexity of extra networks between backbone and decoder. In this paper, we find that the key to these issues is the adaptability of decoders for casting queries to varying objects. Accordingly, we propose a fast-converging query-based detector, named AdaMixer, by improving the adaptability of query-based decoding processes in two aspects. First, each query adaptively samples features over space and scales based on estimated offsets, which allows AdaMixer to efficiently attend to the coherent regions of objects. Then, we dynamically decode these sampled features with an adaptive MLP-Mixer under the guidance of each query. Thanks to these two critical designs, AdaMixer enjoys architectural simplicity without requiring dense attentional encoders or explicit pyramid networks. On the challenging MS COCO benchmark, AdaMixer with ResNet-50 as the backbone, with 12 training epochs, reaches up to 45.0 AP on the validation set along with 27.9 APs in detecting small objects. With the longer training scheme, AdaMixer with ResNeXt-101-DCN and Swin-S reaches 49.5 and 51.3 AP. Our work sheds light on a simple, accurate, and fast converging architecture for query-based object detectors. The code is made available at https://github.com/MCG-NJU/AdaMixer.

Recognition in low quality face datasets is challenging because facial attributes are obscured and degraded. Advances in margin-based loss functions have resulted in enhanced discriminability of faces in the embedding space. Further, previous studies have studied the effect of adaptive losses to assign more importance to misclassified (hard) examples. In this work, we introduce another aspect of adaptiveness in the loss function, namely the image quality. We argue that the strategy to emphasize misclassified samples should be adjusted according to their image quality. Specifically, the relative importance of easy or hard samples should be based on the sample's image quality. We propose a new loss function that emphasizes samples of different difficulties based on their image quality. Our method achieves this in the form of an adaptive margin function by approximating the image quality with feature norms. Extensive experiments show that our method, AdaFace, improves the face recognition performance over the state-of-the-art (SoTA) on four datasets (IJB-B, IJB-C, IJB-S and TinyFace). Code and models are released in Supp.

While methods that regress 3D human meshes from images have progressed rapidly, the estimated body shapes often do not capture the true human shape. This is problematic since, for many applications, accurate body shape is as important as pose. The key reason that body shape accuracy lags pose accuracy is the lack of data. While humans can label 2D joints, and these constrain 3D pose, it is not so easy to "label" 3D body shape. Since paired data with images and 3D body shape are rare, we exploit two sources of information: (1) we collect internet images of diverse "fashion" models together with a small set of anthropometric measurements; (2) we collect linguistic shape attributes for a wide range of 3D body meshes and the model images. Taken together, these datasets provide sufficient constraints to infer dense 3D shape. We exploit the anthropometric measurements and linguistic shape attributes in several novel ways to train a neural network, called SHAPY, that regresses 3D human pose and shape from an RGB image. We evaluate SHAPY on public benchmarks, but note that they either lack significant body shape variation, ground-truth shape, or clothing variation. Thus, we collect a new dataset for evaluating 3D human shape estimation, called HBW, containing photos of "Human Bodies in the Wild" for which we have ground-truth 3D body scans. On this new benchmark, SHAPY significantly outperforms state-of-the-art methods on the task of 3D body shape estimation. This is the first demonstration that 3D body shape regression from images can be trained from easy-to-obtain anthropometric measurements and linguistic shape attributes. Our model and data are available at: shapy.is.tue.mpg.de

This paper is about the old Wahba problem in its more general form, which we call "simultaneous rotation and correspondence search". In this generalization we need to find a rotation that best aligns two partially overlapping 3D point sets, of sizes m and n respectively with m\geq n. We first propose a solver, \texttt ARCS , that i) assumes noiseless point sets in general position, ii) requires only 2 inliers, iii) uses O(m\log m) time and O(m) space, and iv) can successfully solve the problem even with, e.g., m,n~ 10^6 in about 0.1 seconds. We next robustify \texttt ARCS to noise, for which we approximately solve consensus maximization problems using ideas from robust subspace learning and interval stabbing. Thirdly, we refine the approximately found consensus set by a Riemannian subgradient descent approach over the space of unit quaternions, which we show converges globally to an \varepsilon-stationary point in O(\varepsilon^ -4 ) iterations, or locally to the ground-truth at a linear rate in the absence of noise. We combine these algorithms into \texttt ARCS+ , to simultaneously search for rotations and correspondences. Experiments show that \texttt ARCS+ achieves state-of-the-art performance on large-scale datasets with more than 10^6 points with a 10^4 time-speedup over alternative methods. https://github.com/liangzu/ARCS

We describe a method to deal with performance drop in semantic segmentation caused by viewpoint changes within multi-camera systems, where temporally paired images are readily available, but the annotations may only be abundant for a few typical views. Existing methods alleviate performance drop via domain alignment in a shared space and assume that the mapping from the aligned space to the output is transferable. However, the novel content induced by viewpoint changes may nullify such a space for effective alignments, thus resulting in negative adaptation. Our method works without aligning any statistics of the images between the two domains. Instead, it utilizes a novel attention-based view transformation network trained only on color images to hallucinate the semantic images for the target. Despite the lack of supervision, the view transformation network can still generalize to semantic images thanks to the induced "information transport" bias. Furthermore, to resolve ambiguities in converting the semantic images to semantic labels, we treat the view transformation network as a functional representation of an unknown mapping implied by the color images and propose functional label hallucination to generate pseudo-labels with uncertainties in the target domains. Our method surpasses baselines built on state-of-the-art correspondence estimation and view synthesis methods. Moreover, it outperforms the state-of-the-art unsupervised domain adaptation methods that utilize self-training and adversarial domain alignments. Our code and dataset will be made publicly available.

We introduce A-ViT, a method that adaptively adjusts the inference cost of vision transformer ViT for images of different complexity. A-ViT achieves this by automatically reducing the number of tokens in vision transformers that are processed in the network as inference proceeds. We reformulate Adaptive Computation Time (ACT) for this task, extending halting to discard redundant spatial tokens. The appealing architectural properties of vision transformers enables our adaptive token reduction mechanism to speed up inference without modifying the network architecture or inference hardware. We demonstrate that A-ViT requires no extra parameters or sub-network for halting, as we base the learning of adaptive halting on the original network parameters. We further introduce distributional prior regularization that stabilizes training compared to prior ACT approaches. On the image classification task (ImageNet1K), we show that our proposed A-ViT yields high efficacy in filtering informative spatial features and cutting down on the overall compute. The proposed method improves the throughput of DeiT-Tiny by 62% and DeiT-Small by 38% with only 0.3% accuracy drop, outperforming prior art by a large margin.

Deep Neural Networks (DNNs) are known to make overconfident mistakes, which makes their use problematic in safety-critical applications. State-of-the-art (SOTA) calibration techniques improve on the confidence of predicted labels alone, and leave the confidence of non-max classes (e.g. top-2, top-5) uncalibrated. Such calibration is not suitable for label refinement using post-processing. Further, most SOTA techniques learn a few hyper-parameters post-hoc, leaving out the scope for image, or pixel specific calibration. This makes them unsuitable for calibration under domain shift, or for dense prediction tasks like semantic segmentation. In this paper, we argue for intervening at the train time itself, so as to directly produce calibrated DNN models. We propose a novel auxiliary loss function: Multi-class Difference in Confidence and Accuracy (MDCA), to achieve the same. MDCA can be used in conjunction with other application/task specific loss functions. We show that training with MDCA leads to better calibrated models in terms of Expected Calibration Error (ECE), and Static Calibration Error (SCE) on image classification, and segmentation tasks. We report ECE(SCE) score of 0.72 (1.60) on the CIFAR100 dataset, in comparison to 1.90 (1.71) by the SOTA. Under domain shift, a ResNet-18 model trained on PACS dataset using MDCA gives a average ECE(SCE) score of 19.7 (9.7) across all domains, compared to 24.2 (11.8) by the SOTA. For segmentation task, we report a 2x reduction in calibration error on PASCAL-VOC dataset in comparison to Focal Loss. Finally, MDCA training improves calibration even on imbalanced data, and for natural language classification tasks.

While datasets with single-label supervision have propelled rapid advances in image classification, additional annotations are necessary in order to quantitatively assess how models make predictions. To this end, for a subset of ImageNet samples, we collect segmentation masks for the entire object and 18 informative attributes. We call this dataset RIVAL10 (RIch Visual Attributes with Localization), consisting of roughly 26k instances over 10 classes. Using RIVAL10, we evaluate the sensitivity of a broad set of models to noise corruptions in foregrounds, backgrounds and attributes. In our analysis, we consider diverse state-of-the-art architectures (ResNets, Transformers) and training procedures (CLIP, SimCLR, DeiT, Adversarial Training). We find that, somewhat surprisingly, in ResNets, adversarial training makes models more sensitive to the background compared to foreground than standard training. Similarly, contrastively-trained models also have lower relative foreground sensitivity in both transformers and ResNets. Lastly, we observe intriguing adaptive abilities of transformers to increase relative foreground sensitivity as corruption level increases. Using saliency methods, we automatically discover spurious features that drive the background sensitivity of models and assess alignment of saliency maps with foregrounds. Finally, we quantitatively study the attribution problem for neural features by comparing feature saliency with ground-truth localization of semantic attributes.

Observing that the 3D captioning task and the 3D grounding task contain both shared and complementary information in nature, in this work, we propose a unified framework to jointly solve these two distinct but closely related tasks in a synergistic fashion, which consists of both shared task-agnostic modules and lightweight task-specific modules. On one hand, the shared task-agnostic modules aim to learn precise locations of objects, fine-grained attribute features to characterize different objects, and complex relations between objects, which benefit both captioning and visual grounding. On the other hand, by casting each of the two tasks as the proxy task of another one, the lightweight task-specific modules solve the captioning task and the grounding task respectively. Extensive experiments and ablation study on three 3D vision and language datasets demonstrate that our joint training framework achieves significant performance gains for each individual task and finally improves the state-of-the-art performance for both captioning and grounding tasks.

3D visual grounding aims to locate the referred target object in 3D point cloud scenes according to a free-form language description. Previous methods mostly follow a two-stage paradigm, i.e., language-irrelevant detection and cross-modal matching, which is limited by the isolated architecture. In such a paradigm, the detector needs to sample keypoints from raw point clouds due to the inherent properties of 3D point clouds (irregular and large-scale), to generate the corresponding object proposal for each keypoint. However, sparse proposals may leave out the target in detection, while dense proposals may confuse the matching model. Moreover, the language-irrelevant detection stage can only sample a small proportion of keypoints on the target, deteriorating the target prediction. In this paper, we propose a 3D Single-Stage Referred Point Progressive Selection (3D-SPS) method, which progressively selects keypoints with the guidance of language and directly locates the target. Specifically, we propose a Description-aware Keypoint Sampling (DKS) module to coarsely focus on the points of language-relevant objects, which are significant clues for grounding. Besides, we devise a Target-oriented Progressive Mining (TPM) module to finely concentrate on the points of the target, which is enabled by progressive intra-modal relation modeling and inter-modal target mining. 3D-SPS bridges the gap between detection and matching in the 3D visual grounding task, localizing the target at a single stage. Experiments demonstrate that 3D-SPS achieves state-of-the-art performance on both ScanRefer and Nr3D/Sr3D datasets.

Learning a disentangled, interpretable, and structured latent representation in 3D generative models of faces and bodies is still an open problem. The problem is particularly acute when control over identity features is required. In this paper, we propose an intuitive yet effective self-supervised approach to train a 3D shape variational autoencoder (VAE) which encourages a disentangled latent representation of identity features. Curating the mini-batch generation by swapping arbitrary features across different shapes allows to define a loss function leveraging known differences and similarities in the latent representations. Experimental results conducted on 3D meshes show that state-of-the-art methods for latent disentanglement are not able to disentangle identity features of faces and bodies. Our proposed method properly decouples the generation of such features while maintaining good representation and reconstruction capabilities. Our code and pre-trained models are available at github.com/simofoti/3DVAE-SwapDisentangled.

Visual content creation has spurred a soaring interest given its applications in mobile photography and AR / VR. Style transfer and single-image 3D photography as two representative tasks have so far evolved independently. In this paper, we make a connection between the two, and address the challenging task of 3D photo stylization - generating stylized novel views from a single image given an arbitrary style. Our key intuition is that style transfer and view synthesis have to be jointly modeled. To this end, we propose a deep model that learns geometry-aware content features for stylization from a point cloud representation of the scene, resulting in high-quality stylized images that are consistent across views. Further, we introduce a novel training protocol to enable the learning using only 2D images. We demonstrate the superiority of our method via extensive qualitative and quantitative studies, and showcase key applications of our method in light of the growing demand for 3D content creation from 2D image assets.