CVPR 2021 Papers

We aim to infer 3D shape and pose of objects from a single image and propose a learning-based approach that can train from unstructured image collections, using only segmentation outputs from off-the-shelf recognition systems as supervisory signal (i.e. 'shelf-supervised'). We first infer a volumetric representation in a canonical frame, along with the camera pose for the input image. We enforce the representation to be geometrically consistent with both appearance and masks, and also that the synthesized novel views are indistinguishable from image collections. The coarse volumetric prediction is then converted to a mesh-based representation, which is further refined in the predicted camera frame. These two steps allow both shape-pose factorization from unannotated images and reconstruction of per-instance shape in finer details. We report performance on both synthetic and real-world datasets and demonstrate the scalability of our approach on 50 categories in the wild, an order of magnitude more classes than existing works.

Predicting future trajectories of traffic agents in highly interactive environments is an essential and challenging problem for the safe operation of autonomous driving systems. On the basis of the fact that self-driving vehicles are equipped with various types of sensors (e.g., LiDAR scanner, RGB camera, radar, etc.), we propose a Cross-Modal Embedding framework that aims to benefit from the use of multiple input modalities. At training time, our model learns to embed a set of complementary features in a shared latent space by jointly optimizing the objective functions across different types of input data. At test time, a single input modality (e.g., LiDAR data) is required to generate predictions from the input perspective (i.e., in the LiDAR space), while taking advantages from the model trained with multiple sensor modalities. An extensive evaluation is conducted to show the efficacy of the proposed framework using two benchmark driving datasets.

In this paper, we present a technique for estimating the geometry and reflectance of objects using only a camera, flashlight, and optionally a tripod. We propose a simple data capture technique in which the user goes around the object, illuminating it with a flashlight and capturing only a few images. Our main technical contribution is the introduction of a recursive neural architecture, which can predict geometry and reflectance at 2^kx2^k resolution given an input image at 2^kx2^k and estimated geometry and reflectance from the previous step at 2^(k-1)x2^(k-1). This recursive architecture, termed RecNet, is trained with 256x256 resolution but can easily operate on 1024x1024 images during inference. We show that our method produces more accurate surface normal and albedo, especially in regions of specular highlights and cast shadows, compared to previous approaches, given three or fewer input images.

The sky exhibits a unique spatial polarization pattern by scattering the unpolarized sun light. Just like insects use this unique angular pattern to navigate, we use it to map pixels to directions on the sky. That is, we show that the unique polarization pattern encoded in the polarimetric appearance of an object captured under the sky can be decoded to reveal the surface normal at each pixel. We derive a polarimetric reflection model of a diffuse plus mirror surface lit by the sun and a clear sky. This model is used to recover the per-pixel surface normal of an object from a single polarimetric image or from multiple polarimetric images captured under the sky at different times of the day. We experimentally evaluate the accuracy of our shape-from-sky method on a number of real objects of different surface compositions. The results clearly show that this passive approach to fine-geometry recovery that fully leverages the unique illumination made by nature is a viable option for 3D sensing. With the advent of quad-Bayer polarization chips, we believe the implications of our method span a wide range of domains.

Weakly supervised object localization (WSOL) aims to localize objects by only utilizing image-level labels. Class activation maps (CAMs) are the commonly used features to achieve WSOL. However, previous CAM-based methods did not take full advantage of the shallow features, despite their importance for WSOL. Because shallow features are easily buried in background noise through conventional fusion. In this paper, we propose a simple but effective Shallow feature-aware Pseudo supervised Object Localization (SPOL) model for accurate WSOL, which makes the utmost of low-level features embedded in shallow layers. In practice, our SPOL model first generates the CAMs through a novel element-wise multiplication of shallow and deep feature maps, which filters the background noise and generates sharper boundaries robustly. Besides, we further propose a general class-agnostic segmentation model to achieve the accurate object mask, by only using the initial CAMs as the pseudo label without any extra annotation. Eventually, a bounding box extractor is applied to the object mask to locate the target. Experiments verify that our SPOL outperforms the state-of-the-art on both CUB-200 and ImageNet-1K benchmarks, achieving 93.44% and 67.15% (i.e., 3.93% and 2.13% improvement) Top-5 localization accuracy, respectively.

Perhaps surprisingly sewerage infrastructure is one of the most costly infrastructures in modern society. Sewer pipes are manually inspected to determine whether the pipes are defective. However, this process is limited by the number of qualified inspectors and the time it takes to inspect a pipe. Automatization of this process is therefore of high interest. So far, the success of computer vision approaches for sewer defect classification has been limited when compared to the success in other fields mainly due to the lack of public datasets. To this end, in this work we present a large novel and publicly available multi-label classification dataset for image-based sewer defect classification called Sewer-ML. The Sewer-ML dataset consists of 1.3 million images annotated by professional sewer inspectors from three different utility companies across nine years. Together with the dataset, we also present a benchmark algorithm and a novel metric for assessing performance. The benchmark algorithm is a result of evaluating 12 state-of-the-art algorithms, six from the sewer defect classification domain and six from the multi-label classification domain, and combining the best performing algorithms. The novel metric is a class-importance weighted F2 score, F2-CIW, reflecting the economic impact of each class, used together with the normal pipe F1 score, F1-Normal. The benchmark algorithm achieves an F2-CIW score of 55.11% and F1-Normal score of 90.94%, leaving ample room for improvement on the Sewer-ML dataset. The code, models, and dataset are available at the project page http://vap.aau.dk/sewer-ml

Generative modeling of set-structured data, such as point clouds, requires reasoning over local and global structures at various scales. However, adopting multi-scale frameworks for ordinary sequential data to a set-structured data is nontrivial as it should be invariant to the permutation of its elements. In this paper, we propose SetVAE, a hierarchical variational autoencoder for sets. Motivated by recent progress in set encoding, we build SetVAE upon attentive modules that first partition the set and project the partition back to the original cardinality. Exploiting this module, our hierarchical VAE learns latent variables at multiple scales, capturing coarse-to-fine dependency of the set elements while achieving permutation invariance. We evaluate our model on point cloud generation task and achieve competitive performance to the prior arts with substantially smaller model capacity. We qualitatively demonstrate that our model generalizes to unseen set sizes and learns interesting subset relations without supervision. Our implementation is available at https://github.com/jw9730/setvae.

We propose a novel pool-based Active Learning frame-work constructed on a sequential Graph Convolution Net-work (GCN). Each image's feature from a pool of data rep-resents a node in the graph and the edges encode their similarities. With a small number of randomly sampled images as seed labelled examples, we learn the parameters of the graph to distinguish labelled vs unlabelled nodes by minimising the binary cross-entropy loss. GCN performs message-passing operations between the nodes, and hence, induces similar representations of the strongly associated nodes. We exploit these characteristics of GCN to select the unlabelled examples which are sufficiently different from la-belled ones. To this end, we utilise the graph node embed-dings and their confidence scores and adapt sampling techniques such as CoreSet and uncertainty-based methods to query the nodes. We flip the label of newly queried nodes from unlabelled to labelled, re-train the learner to optimise the downstream task and the graph to minimise its modified objective. We continue this process within a fixed budget. We evaluate our method on 6 different benchmarks: 4 real image classification, 1 depth-based hand pose estimation and 1 synthetic RGB image classification datasets. Our method outperforms several competitive baselines such as VAAL, Learning Loss, CoreSet and attains the new state-of-the-art performance on multiple applications.

We propose a framework for sequence-to-sequence contrastive learning (SeqCLR) of visual representations, which we apply to text recognition. To account for the sequence-to-sequence structure, each feature map is divided into different instances over which the contrastive loss is computed. This operation enables us to contrast in a sub-word level, where from each image we extract several positive pairs and multiple negative examples. To yield effective visual representations for text recognition, we further suggest novel augmentation heuristics, different encoder architectures and custom projection heads. Experiments on handwritten text and on scene text show that when a text decoder is trained on the learned representations, our method outperforms non-sequential contrastive methods. In addition, when the amount of supervision is reduced, SeqCLR significantly improves performance compared with supervised training, and when fine-tuned with 100% of the labels, our method achieves state-of-the-art results on standard handwritten text recognition benchmarks.

Generalization to out-of-distribution data has been a problem for Visual Question Answering (VQA) models. To measure generalization to novel questions, we propose to separate them into "skills" and "concepts". "Skills" are visual tasks, such as counting or attribute recognition, and are applied to "concepts" mentioned in the question, such as objects and people. VQA methods should be able to compose skills and concepts in novel ways, regardless of whether the specific composition has been seen in training, yet we demonstrate that existing models have much to improve upon towards handling new compositions. We present a novel method for learning to compose skills and concepts that separates these two factors implicitly within a model by learning grounded concept representations and disentangling the encoding of skills from that of concepts. We enforce these properties with a novel contrastive learning procedure that does not rely on external annotations and can be learned from unlabeled image-question pairs. Experiments demonstrate the effectiveness of our approach for improving compositional and grounding performance.

While deep learning (DL)-based video deraining methods have achieved significant success recently, they still exist two major drawbacks. Firstly, most of them do not sufficiently model the characteristics of rain layers of rainy videos. In fact, the rain layers exhibit strong physical properties (e.g., direction, scale and thickness) in spatial dimension and natural continuities in temporal dimension, and thus can be generally modelled by the spatial-temporal process in statistics. Secondly, current DL-based methods seriously depend on the labeled synthetic training data, whose rain types are always deviated from those in unlabeled real data. Such gap between synthetic and real data sets leads to poor performance when applying them in real scenarios. Against these issues, this paper proposes a new semisupervised video deraining method, in which a dynamic rain generator is employed to fit the rain layer, expecting to better depict its insightful characteristics. Specifically, such dynamic generator consists of one emission model and one transition model to simultaneously encode the spatially physical structure and temporally continuous changes of rain streaks, respectively, which both are parameterized as deep neural networks (DNNs). Further more, different prior formats are designed for the labeled synthetic and unlabeled real data, so as to fully exploit the common knowledge underlying them. Last but not least, we also design a Monte Carlo EM algorithm to solve this model. Extensive experiments are conducted to verify the superiorities of the proposed semi-supervised deraining model.

We introduce a novel approach to generate diverse high fidelity texture maps for 3D human meshes in a semi-supervised setup. Given a segmentation mask defining the layout of the semantic regions in the texture map, our network generates high-resolution textures with a variety of styles, that are then used for rendering purposes. To accomplish this task, we propose a Region-adaptive Adversarial Variational AutoEncoder (ReAVAE) that learns the probability distribution of the style of each region individually so that the style of the generated texture can be controlled by sampling from the region-specific distributions. In addition, we introduce a data generation technique to augment our training set with data lifted from single-view RGB inputs. Our training strategy allows the mixing of reference image styles with arbitrary styles for different regions, a property which can be valuable for virtual try-on AR/VR applications. Experimental results show that our method synthesizes better texture maps compared to prior work while enabling independent layout and style controllability.

Semantic segmentation has made tremendous progress in recent years. However, satisfying performance highly depends on a large number of pixel-level annotations. Therefore, in this paper, we focus on the semi-supervised segmentation problem where only a small set of labeled data is provided with a much larger collection of totally unlabeled images. Nevertheless, due to the limited annotations, models may overly rely on the contexts available in the training data, which causes poor generalization to the scenes unseen before. A preferred high-level representation should capture the contextual information while not losing self-awareness. Therefore, we propose to maintain the context-aware consistency between features of the same identity but with different contexts, making the representations robust to the varying environments. Moreover, we present the Directional Contrastive Loss (DC Loss) to accomplish the consistency in a pixel-to-pixel manner, only requiring the feature with lower quality to be aligned towards its counterpart. In addition, to avoid the false-negative samples and filter the uncertain positive samples, we put forward two sampling strategies. Extensive experiments show that our simple yet effective method surpasses current state-of-the-art methods by a large margin and also generalizes well with extra image-level annotations.

In this paper, we study the semi-supervised semantic segmentation problem via exploring both labeled data and extra unlabeled data. We propose a novel consistency regularization approach, called cross pseudo supervision (CPS). Our approach imposes the consistency on two segmentation networks perturbed with different initialization for the same input image. The pseudo one-hot label map, output from one perturbed segmentation network, is used to supervise the other segmentation network with the standard cross-entropy loss, and vice versa. The CPS consistency has two roles: encourage high similarity between the predictions of two perturbed networks for the same input image, and expand training data by using the unlabeled data with pseudo labels.

Data-driven based approaches, in spite of great success in many tasks, have poor generalization when applied to unseen image domains, and require expensive cost of annotation especially for dense pixel prediction tasks such as semantic segmentation. Recently, both unsupervised domain adaptation (UDA) from large amounts of synthetic data and semi-supervised learning (SSL) with small set of labeled data have been studied to alleviate this issue. However, there is still a large gap on performance compared to their supervised counterparts. We focus on a more practical setting of semi-supervised domain adaptation (SSDA) where both a small set of labeled target data and large amounts of labeled source data are available. To address the task of SSDA, a novel framework based on dual-level domain mixing is proposed. The proposed framework consists of three stages. First, two kinds of data mixing methods are proposed to reduce domain gap in both region-level and sample-level respectively. We can obtain two complementary domain-mixed teachers based on dual-level mixed data from holistic and partial views respectively. Then, a student model is learned by distilling knowledge from these two teachers. Finally, pseudo labels of unlabeled data are generated in a self-training manner for another few rounds of teachers training. Extensive experimental results have demonstrated the effectiveness of our proposed framework on synthetic-to-real semantic segmentation benchmarks.

Learning to recognize actions from only a handful of labeled videos is a challenging problem due to the scarcity of tediously collected activity labels. We approach this problem by learning a two-pathway temporal contrastive model using unlabeled videos at two different speeds leveraging the fact that changing video speed does not change an action. Specifically, we propose to maximize the similarity between encoded representations of the same video at two different speeds as well as minimize the similarity between different videos played at different speeds. This way we use the rich supervisory information in terms of `time' that is present in otherwise unsupervised pool of videos. With this simple yet effective strategy of manipulating video playback rates, we considerably outperform video extensions of sophisticated state-of-the-art semi-supervised image recognition methods across multiple diverse benchmark datasets and network architectures. Interestingly, our proposed approach benefits from out-of-domain unlabeled videos showing generalization and robustness. We also perform rigorous ablations and analysis to validate our approach. Project page: https://cvir.github.io/TCL/.

Estimating 3D hand and object pose from a single image is an extremely challenging problem: hands and objects are often self-occluded during interactions, and the 3D annotations are scarce as even humans cannot directly label the ground-truths from a single image perfectly. To tackle these challenges, we propose a unified framework for estimating the 3D hand and object poses with semi-supervised learning. We build a joint learning framework where we perform explicit contextual reasoning between hand and object representations. Going beyond limited 3D annotations in a single image, we leverage the spatial-temporal consistency in large-scale hand-object videos as a constraint for generating pseudo labels in semi-supervised learning. Our method not only improves hand pose estimation in challenging real-world dataset, but also substantially improve the object pose which has fewer ground-truths per instance. By training with large-scale diverse videos, our model also generalizes better across multiple out-of-domain datasets. Project page and code: https://stevenlsw.github.io/Semi-Hand-Object

Most existing video text detection methods track texts with appearance features, which are easily influenced by the change of perspective and illumination. Compared with appearance features, semantic features are more robust cues for matching text instances. In this paper, we propose an end-to-end trainable video text detector that tracks texts based on semantic features. First, we introduce a new character center segmentation branch to extract semantic features, which encode the category and position of characters. Then we propose a novel appearance-semantic-geometry descriptor to track text instances, in which semantic features can improve the robustness against appearance changes. To overcome the lack of character-level annotations, we propose a novel weakly-supervised character center detection module, which only uses word-level annotated real images to generate character-level labels. The proposed method achieves state-of-the-art performance on three video text benchmarks ICDAR 2013 Video, Minetto and RT-1K, and two Chinese scene text benchmarks CASIA10K and MSRA-TD500.

Few-shot class incremental learning (FSCIL) portrays the problem of learning new concepts gradually, where only a few examples per concept are available to the learner. Due to the limited number of examples for training, the techniques developed for standard incremental learning cannot be applied verbatim to FSCIL. In this work, we introduce a distillation algorithm to address the problem of FSCIL and propose to make use of semantic information during training. To this end, we make use of word embeddings as semantic information which is cheap to obtain and which facilitate the distillation process. Furthermore, we propose a method based on an attention mechanism on multiple parallel embeddings of visual data to align visual and semantic vectors, which reduces issues related to catastrophic forgetting. Via experiments on MiniImageNet, CUB200, and CIFAR100 dataset, we establish new state-of-the-art results by outperforming existing approaches.