CVPR 2021 Papers

View synthesis is usually done by an autoencoder, in which the encoder maps a source view image into a latent content code, and the decoder transforms it into a target view image according to the condition. However, the source contents are often not well kept in this setting, which leads to unnecessary changes during the view translation. Although adding skipped connections, like Unet, alleviates the problem, but it often causes the failure on the view conformity. This paper proposes a new architecture by performing the source-to-target deformation in an iterative way. Instead of simply incorporating the features from multiple layers of the encoder, we design soft and hard deformation modules, which warp the encoder features to the target view at different resolutions, and give results to the decoder to complement the details. Particularly, the current warping flow is not only used to align the feature of the same resolution, but also as an approximation to coarsely deform the high resolution feature. Then the residual flow is estimated and applied in the high resolution, so that the deformation is built up in the coarse-to-fine fashion. To better constrain the model, we synthesize a rough target view image based on the intermediate flows and their warped features. The extensive ablation studies and the final results on two different data sets show the effectiveness of the proposed model.

We present a method that synthesizes novel views of complex scenes by interpolating a sparse set of nearby views. The core of our method is a network architecture that includes a multilayer perceptron and a ray transformer that estimates radiance and volume density at continuous 5D locations (3D spatial locations and 2D viewing directions), drawing appearance information on the fly from multiple source views. By drawing on source views at render time, our method hearkens back to classic work on image-based rendering (IBR), and allows us to render high-resolution imagery. Unlike neural scene representation work that optimizes per-scene functions for rendering, we learn a generic view interpolation function that generalizes to novel scenes. We render images using classic volume rendering, which is fully differentiable and allows us to train using only multi-view posed images as supervision. Experiments show that our method outperforms recent novel view synthesis methods that also seek to generalize to novel scenes. Further, if fine-tuned on each scene, our method is competitive with state-of-the-art single-scene neural rendering methods.

Recent works on two-stage cross-domain detection have widely explored the local feature patterns to achieve more accurate adaptation results. These methods heavily rely on the region proposal mechanisms and ROI-based instance-level features to design fine-grained feature alignment modules with respect to the foreground objects. However, for one-stage detectors, it is hard or even impossible to obtain explicit instance-level features in the detection pipelines. Motivated by this, we propose an Implicit Instance-Invariant Network (I3Net), which is tailored for adapting one-stage detectors and implicitly learns instance-invariant features via exploiting the natural characteristics of deep features in different layers. Specifically, we facilitate the adaptation from three aspects: (1) Dynamic and Class-Balanced Reweighting (DCBR) strategy, which considers the coexistence of intra-domain and intra-class variations to assign larger weights to those sample-scarce categories and easy-to-adapt samples; (2) Category-aware Object Pattern Matching (COPM) module, which boosts the cross-domain foreground objects matching guided by the categorical information and suppresses the uninformative background features; (3) Regularized Joint Category Alignment (RJCA) module, which jointly enforces the category alignment at different domain-specific layers with a consistency regularization. Experiments reveal that I3Net exceeds the state-of-the-art performance on benchmark datasets.

Image and video descriptors are an omnipresent tool in computer vision and its application fields like mobile robotics. Many hand-crafted and in particular learned image descriptors are numerical vectors with a potentially (very) large number of dimensions. Practical considerations like memory consumption or time for comparisons call for the creation of compact representations. In this paper, we use hyperdimensional computing (HDC) as an approach to systematically combine information from a set of vectors in a single vector of the same dimensionality. HDC is a known technique to perform symbolic processing with distributed representations in numerical vectors with thousands of dimensions. We present a HDC implementation that is suitable for processing the output of existing and future (deep learning based) image descriptors. We discuss how this can be used as a framework to process descriptors together with additional knowledge by simple and fast vector operations. A concrete outcome is a novel HDC-based approach to aggregate a set of local image descriptors together with their image positions in a single holistic descriptor. The comparison to available holistic descriptors and aggregation methods on a series of standard mobile robotics place recognition experiments shows a 20% improvement in average performance and >2x better worst-case performance compared to runner-up.

We present a novel, real-time, semantic segmentation network in which the encoder both encodes and generates the parameters (weights) of the decoder. Furthermore, to allow maximal adaptivity, the weights at each decoder block vary spatially. For this purpose, we design a new type of hypernetwork, composed of a nested U-Net for drawing higher level context features, a multi-headed weight generating module which generates the weights of each block in the decoder immediately before they are consumed, for efficient memory utilization, and a primary network that is composed of novel dynamic patch-wise convolutions. Despite the usage of less-conventional blocks, our architecture obtains real-time performance. In terms of the runtime vs. accuracy trade-off, we surpass state of the art (SotA) results on popular semantic segmentation benchmarks: PASCAL VOC 2012 (val. set) and real-time semantic segmentation on Cityscapes, and CamVid. The code is available: https://nirkin.com/hyperseg.

Deep neural networks are susceptible to catastrophic forgetting: when encountering a new task, they can only remember the new task and fail to preserve its ability to accomplish previously learned tasks. In this paper, we study the problem of lifelong learning for generative models and propose a novel and generic continual learning framework Hyper-LifelongGAN which is more scalable compared with state-of-the-art approaches. Given a sequence of tasks, the conventional convolutional filters are factorized into the dynamic base filters which are generated using task specific filter generators, and deterministic weight matrix which linearly combines the base filters and is shared across different tasks. Moreover, the shared weight matrix is multiplied by task specific coefficients to introduce more flexibility in combining task specific base filters differently for different tasks. Attributed to the novel architecture, the proposed method can preserve or even improve the generation quality at a low cost of parameters. We validate Hyper-LifelongGAN on diverse image-conditioned generation tasks, extensive ablation studies and comparisons with state-of-the-art models are carried out to show that the proposed approach can address catastrophic forgetting effectively.

We address rotation averaging (RA) and its application to real-world 3D reconstruction. Local optimisation based approaches are the de facto choice, though they only guarantee a local optimum. Global optimisers ensure global optimality in low noise conditions, but they are inefficient and may easily deviate under the influence of outliers or elevated noise levels. We push the envelope of rotation averaging by leveraging the advantages of a global RA method and a local RA method. Combined with a fast view graph filtering as preprocessing, the proposed hybrid approach is robust to outliers. We further apply the proposed hybrid rotation averaging approach to incremental Structure from Motion (SfM), the accuracy and robustness of SfM are both improved by adding the resulting global rotations as regularisers to bundle adjustment. Overall, we demonstrate high practicality of the proposed method as bad camera poses are effectively corrected and drift is reduced.

Message passing neural network has been an effective method to represent dependencies among nodes by propagating messages. However, most of message passing algorithms focus on one structure and the messages are estimated by one single approach. For the real-world data, like facial action units (AUs), the dependencies may vary in terms of different expressions and individuals. In this paper, we propose a novel hybrid message passing neural network with performance-driven structures (HMP-PS), which combines complementary message passing methods and captures more possible structures in a Bayesian manner. Particularly, a performance-driven Monte Carlo Markov Chain sampling method is proposed for generating high performance graph structures. Besides, the hybrid message passing is proposed to combine different types of messages, which provide the complementary information. The contribution of each type of message is adaptively adjusted along with different inputs. The experiments on two widely used benchmark datasets, i.e., BP4D and DISFA, validate that our proposed method can achieve the state-of-the-art performance.

Model-based 3D pose and shape estimation methods reconstruct a full 3D mesh for the human body by estimating several parameters. However, learning the abstract parameters is a highly non-linear process and suffers from image-model misalignment, leading to mediocre model performance. In contrast, 3D keypoint estimation methods combine deep CNN network with the volumetric representation to achieve pixel-level localization accuracy but may predict unrealistic body structure. In this paper, we address the above issues by bridging the gap between body mesh estimation and 3D keypoint estimation. We propose a novel hybrid inverse kinematics solution (HybrIK). HybrIK directly transforms accurate 3D joints to relative body-part rotations for 3D body mesh reconstruction, via the twist-and-swing decomposition. The swing rotation is analytically solved with 3D joints, and the twist rotation is derived from the visual cues through the neural network. We show that HybrIK preserves both the accuracy of 3D pose and the realistic body structure of the parametric human model, leading to a pixel-aligned 3D body mesh and a more accurate 3D pose than the pure 3D keypoint estimation methods. Without bells and whistles, the proposed method surpasses the state-of-the-art methods by a large margin on various 3D human pose and shape benchmarks. As an illustrative example, HybrIK outperforms all the previous methods by 13.2 mm MPJPE and 21.9 mm PVE on 3DPW dataset. Our code is available at https://github.com/Jeff-sjtu/HybrIK.

We propose a semi-supervised approach for contemporary object detectors following the teacher-student dual model framework. Our method is featured with 1) the exponential moving averaging strategy to update the teacher from the student online, 2) using plenty of region proposals and soft pseudo-labels as the student's training targets, and 3) a light-weighted detection-specific data ensemble for the teacher to generate more reliable pseudo labels. Compared to the recent state-of-the-art - STAC, which uses hard labels on sparsely selected hard pseudo samples, the teacher in our model exposes richer information to the student with soft-labels on many proposals. Our model achieves COCO-style AP of 53.04% on VOC07 val set, 8.4% better than STAC, when using VOC12 as unlabeled data. On MS-COCO, it outperforms prior work when only a small percentage of data is taken as labeled. It also reaches 53.8% AP on MS-COCO test-dev with 3.1% gain over the fully supervised ResNet-152 cascaded R-CNN, by tapping into unlabeled data of a similar size to the labeled data.

In this paper, we address the problem of building pixel-wise dense correspondences between human images under arbitrary camera viewpoints and body poses. Previous methods either assume small motions or rely on discriminative descriptors extracted from local patches, which cannot handle large motion or visually ambiguous body parts, e.g. left v.s. right hand. In contrast, we propose a deep learning framework that maps each pixel to a feature space, where the feature distances reflect the geodesic distances among pixels as if they were projected onto the surface of 3D human scans. To this end, we introduce novel loss functions to push features apart according to their geodesic distances on the surface inside and across images. Without any semantic annotation, the features automatically learn to differentiate visually similar parts and align different subjects into a unified feature space. Extensive experiments show that the learned features can produce accurate correspondences between images with remarkable generalization capabilities on both intra and inter subjects. We demonstrate the effectiveness of our method on a variety of applications such as optical flow, non-rigid tracking, occlusions detection, and human dense pose regression.

Controllable Image Captioning (CIC) -- generating image descriptions following designated control signals -- has received unprecedented attention over the last few years. To emulate the human ability in controlling caption generation, current CIC studies focus exclusively on control signals concerning objective properties, such as contents of interest or descriptive patterns. However, we argue that almost all existing objective control signals have overlooked two indispensable characteristics of an ideal control signal: 1) Event-compatible: all visual contents referred to in a single sentence should be compatible with the described activity. 2) Sample-suitable: the control signals should be suitable for a specific image sample. To this end, we propose a new control signal for CIC: Verb-specific Semantic Roles (VSR). VSR consists of a verb and some semantic roles, which represents a targeted activity and the roles of entities involved in this activity. Given a designated VSR, we first train a grounded semantic role labeling (GSRL) model to identify and ground all entities for each role. Then, we propose a semantic structure planner (SSP) to learn human-like descriptive semantic structures. Lastly, we use a role-shift captioning model to generate the captions. Extensive experiments and ablations demonstrate that our framework can achieve better controllability than several strong baselines on two challenging CIC benchmarks. Besides, we can generate multi-level diverse captions easily. The code is available at: https://github.com/mad-red/VSR-guided-CIC.

We introduce (HPS) Human POSEitioning System, a method to recover the full 3D pose of a human registered with a 3D scan of the surrounding environment using wearable sensors. Using IMUs attached at the body limbs and a head mounted camera looking outwards, HPS fuses camera based self-localization with IMU-based human body tracking. The former provides drift-free but noisy position and orientation estimates while the latter is accurate in the short-term but subject to drift over longer periods of time. We show that our optimization-based integration exploits the benefits of the two, resulting in pose accuracy free of drift. Furthermore, we integrate 3D scene constraints into our optimization, such as foot contact with the ground, resulting in physically plausible motion. HPS complements more common third-person-based 3D pose estimation methods. It allows capturing larger recording volumes and longer periods of motion, and could be used for VR/AR applications where humans interact with the scene without requiring direct line of sight with an external camera, or to train agents that navigate and interact with the environment based on first-person visual input, like real humans. With HPS, we recorded a dataset of humans interacting with large 3D scenes (300-1000 sq.m) consisting of 7 subjects and more than 3 hours of diverse motion. The dataset, code and video will be available on the project page: http://virtualhumans.mpi-inf.mpg.de/hps/.

In this paper, we tackle the problem of human de-occlusion which reasons about occluded segmentation masks and invisible appearance content of humans. In particular, a two-stage framework is proposed to estimate the invisible portions and recover the content inside. For the stage of mask completion, a stacked network structure is devised to refine inaccurate masks from a general instance segmentation model and predict integrated masks simultaneously. Additionally, the guidance from human parsing and typical pose masks are leveraged to bring prior information. For the stage of content recovery, a novel parsing guided attention module is applied to isolate body parts and capture context information across multiple scales. Besides, an Amodal Human Perception dataset (AHP) is collected to settle the task of human de-occlusion. AHP has advantages of providing annotations from real-world scenes and the number of humans is comparatively larger than other amodal perception datasets. Based on this dataset, experiments demonstrate that our method performs over the state-of-the-art techniques in both tasks of mask completion and content recovery. Our AHP dataset is available at https://sydney0zq.github.io/ahp/.

One of the factors that have hindered progress in the areas of sign language recognition, translation, and production is the absence of large annotated datasets. Towards this end, we introduce How2Sign, a multimodal and multiview continuous American Sign Language (ASL) dataset, consisting of a parallel corpus of more than 80 hours of sign language videos and a set of corresponding modalities including speech, English transcripts, and depth. A three-hour subset was further recorded in the Panoptic studio enabling detailed 3D pose estimation. To evaluate the potential of How2Sign for real-world impact, we conduct a study with ASL signers and show that synthesized videos using our dataset can indeed be understood. The study further gives insights on challenges that computer vision should address in order to make progress in this field. Dataset website: http://how2sign.github.io/

Self-supervised visual representation learning has seen huge progress recently, but no large scale evaluation has compared the many models now available. We evaluate the transfer performance of 13 top self-supervised models on 40 downstream tasks, including many-shot and few-shot recognition, object detection, and dense prediction. We compare their performance to a supervised baseline and show that on most tasks the best self-supervised models outperform supervision, confirming the recently observed trend in the literature. We find ImageNet Top-1 accuracy to be highly correlated with transfer to many-shot recognition, but increasingly less so for few-shot, object detection and dense prediction. No single self-supervised method dominates overall, suggesting that universal pre-training is still unsolved. Our analysis of features suggests that top self-supervised learners fail to preserve colour information as well as supervised alternatives, but tend to induce better classifier calibration, and less attentive overfitting than supervised learners.

Since its inception, Visual Question Answering (VQA) is notoriously known as a task, where models are prone to exploit biases in datasets to find shortcuts instead of performing high-level reasoning. Classical methods address this by removing biases from training data, or adding branches to models to detect and remove biases. In this paper, we argue that uncertainty in vision is a dominating factor preventing the successful learning of reasoning in vision and language problems. We train a visual oracle and in a large scale study provide experimental evidence that it is much less prone to exploiting spurious dataset biases compared to standard models. We propose to study the attention mechanisms at work in the visual oracle and compare them with a SOTA Transformer-based model. We provide an in-depth analysis and visualizations of reasoning patterns obtained with an online visualization tool which we make publicly available (https://reasoningpatterns.github.io). We exploit these insights by transferring reasoning patterns from the oracle to a SOTA Transformer-based VQA model taking standard noisy visual inputs via fine-tuning. In experiments we report higher overall accuracy, as well as accuracy on infrequent answers for each question type, which provides evidence for improved generalization and a decrease of the dependency on dataset biases.

Lossy image compression is often limited by the simplicity of the chosen loss measure. Recent research suggests that generative adversarial networks have the ability to overcome this limitation and serve as a multi-modal loss, especially for textures. Together with learned image compression, these two techniques can be used to great effect when relaxing the commonly employed tight measures of distortion. However, convolutional neural network-based algorithms have a large computational footprint. Ideally, an existing conventional codec should stay in place, ensuring faster adoption and adherence to a balanced computational envelope. As a possible avenue to this goal, we propose and investigate how learned image coding can be used as a surrogate to optimise an image for encoding. A learned filter alters the image to optimise a different performance measure or a particular task. Extending this idea with a generative adversarial network, we show how entire textures are replaced by ones that are less costly to encode but preserve a sense of detail. Our approach can remodel a conventional codec to adjust for the MS-SSIM distortion with over 20% rate improvement without any decoding overhead. On task-aware image compression, we perform favourably against a similar but codec-specific approach.

Predictions of certifiably robust classifiers remain constant in a neighborhood of a point, making them resilient to test-time attacks with a guarantee. In this work, we present a previously unrecognized threat to robust machine learning models that highlights the importance of training-data quality in achieving high certified adversarial robustness. Specifically, we propose a novel bilevel optimization-based data poisoning attack that degrades the robustness guarantees of certifiably robust classifiers. Unlike other poisoning attacks that reduce the accuracy of the poisoned models on a small set of target points, our attack reduces the average certified radius (ACR) of an entire target class in the dataset. Moreover, our attack is effective even when the victim trains the models from scratch using state-of-the-art robust training methods such as Gaussian data augmentation [??], MACER [??], and SmoothAdv [??] that achieve high certified adversarial robustness. To make the attack harder to detect, we use clean-label poisoning points with imperceptible distortions. The effectiveness of the proposed method is evaluated by poisoning MNIST and CIFAR10 datasets and training deep neural networks using previously mentioned training methods and certifying the robustness with randomized smoothing. The ACR of the target class, for models trained on generated poison data, can be reduced by more than 30%. Moreover, the poisoned data is transferable to models trained with different training methods and models with different architectures.