ICCV 2021 Papers

In open set recognition, a classifier has to detect unknown classes that are not known at training time. In order to recognize new categories, the classifier has to project the input samples of known classes in very compact and separated regions of the features space for discriminating samples of unknown classes. Recently proposed Capsule Networks have shown to outperform alternatives in many fields, particularly in image recognition, however they have not been fully applied yet to open-set recognition. In capsule networks, scalar neurons are replaced by capsule vectors or matrices, whose entries represent different properties of objects. In our proposal, during training, capsules features of the same known class are encouraged to match a pre-defined gaussian, one for each class. To this end, we use the variational autoencoder framework, with a set of gaussian priors as the approximation for the posterior distribution. In this way, we are able to control the compactness of the features of the same class around the center of the gaussians, thus controlling the ability of the classifier in detecting samples from unknown classes. We conducted several experiments and ablation of our model, obtaining state of the art results on different datasets in the open set recognition and unknown detection tasks.

In click-based interactive segmentation, the mask extraction process is dictated by positive/negative user clicks; however, most existing methods do not fully exploit the user cues, requiring excessive numbers of clicks for satisfactory results. We propose Conditional Diffusion Network(CDNet), which propagates labeled representations from clicks to conditioned destinations with two levels of affinities: Feature Diffusion Module (FDM) spreads features from clicks to potential target regions with global similarity; Pixel Diffusion Module (PDM) diffuses the predicted logits of clicks within locally connected regions. Thus, the information inferred by user clicks could be generalized to proper destinations. In addition, we put forward Diversified Training(DT), which reduces the optimization ambiguity caused by click simulation. With FDM,PDM and DT, CDNet could better understand user's intentions and make better predictions with limited interactions. CDNet achieves state-of-the-art performance on several benchmarks.

The recently-developed DETR approach applies the transformer encoder and decoder architecture to object detection and achieves promising performance. In this paper, we handle the critical issue, slow training convergence, and present a conditional cross-attention mechanism for fast DETR training. Our approach is motivated by that the cross-attention in DETR relies highly on the content embeddings for localizing the four extremities and predicting the box, which increases the need for high-quality content embeddings and thus the training difficulty. Our approach, named conditional DETR, learns a conditional spatial query from the decoder embedding for decoder multi-head cross-attention. The benefit is that through the conditional spatial query, each cross-attention head is able to attend to a band containing a distinct region, e.g., one object extremity or a region inside the object box. This narrows down the spatial range for localizing the distinct regions for object classification and box regression, thus relaxing the dependence on the content embeddings and easing the training. Empirical results show that conditional DETR converges 6.7x faster for the backbones R50 and R101 and 10x faster for stronger backbones DC5-R50 and DC5-R101. Code is available at https://github.com/Atten4Vis/ConditionalDETR.

Video is complex due to large variations in motion and rich content in fine-grained visual details. Abstracting useful information from such information-intensive media requires exhaustive computing resources. This paper studies a two-step alternative that first condenses the video sequence to an informative "frame" and then exploits off-the-shelf image recognition system on the synthetic frame. A valid question is how to define "useful information" and then distill it from a video sequence down to one synthetic frame. This paper presents a novel Informative Frame Synthesis (IFS) architecture that incorporates three objective tasks, i.e., appearance reconstruction, video categorization, motion estimation, and two regularizers, i.e., adversarial learning, color consistency. Each task equips the synthetic frame with one ability, while each regularizer enhances its visual quality. With these, by jointly learning the frame synthesis in an end-to-end manner, the generated frame is expected to encapsulate the required spatio-temporal information useful for video analysis. Extensive experiments are conducted on the large-scale Kinetics dataset. When comparing to baseline methods that map video sequence to a single image, IFS shows superior performance. More remarkably, IFS consistently demonstrates evident improvements on image-based 2D networks and clip-based 3D networks, and achieves comparable performance with the state-of-the-art methods with less computational cost.

Modern deep-learning-based lane detection methods are successful in most scenarios but struggling for lane lines with complex topologies. In this work, we propose CondLaneNet, a novel top-to-down lane detection framework that detects the lane instances first and then dynamically predicts the line shape for each instance. Aiming to resolve lane instance-level discrimination problem, we introduce a conditional lane detection strategy based on conditional convolution and row-wise formulation. Further, we design the Recurrent Instance Module(RIM) to overcome the problem of detecting lane lines with complex topologies such as dense lines and fork lines. Benefit from the end-to-end pipeline which requires little post-process, our method has real-time efficiency. We extensively evaluate our method on three benchmarks of lane detection. Results show that our method achieves state-of-the-art performance on all three benchmark datasets. Moreover, our method has the coexistence of accuracy and efficiency, e.g. a 78.14 F1 score and 220 FPS on CULane. Our code is available at https://github.com/aliyun/ conditional-lane-detection.

Measuring concept generalization, i.e., the extent to which models trained on a set of (seen) visual concepts can be leveraged to recognize a new set of (unseen) concepts, is a popular way of evaluating visual representations, especially in a self-supervised learning framework. Nonetheless, the choice of unseen concepts for such an evaluation is usually made arbitrarily, and independently from the seen concepts used to train representations, thus ignoring any semantic relationships between the two. In this paper, we argue that the semantic relationships between seen and unseen concepts affect generalization performance and propose ImageNet-CoG, a novel benchmark on the ImageNet-21K (IN-21K) dataset that enables measuring concept generalization in a principled way. Our benchmark leverages expert knowledge that comes from WordNet in order to define a sequence of unseen IN-21K concept sets that are semantically more and more distant from the ImageNet-1K (IN-1K) subset, a ubiquitous training set. This allows us to benchmark visual representations learned on IN-1K out-of-the box. We conduct a large-scale study encompassing 31 convolution and transformer-based models and show how different architectures, levels of supervision, regularization techniques and use of web data impact the concept generalization performance.

Despite exciting progress in pre-training for visual-linguistic (VL) representations, very few aspire to a small VL model. In this paper, we study knowledge distillation(KD) to effectively compress a transformer-based large VL model into a small VL model. The major challenge arises from the inconsistent regional visual tokens extracted from different detectors of Teacher and Student, resulting in the misalignment of hidden representations and attention distributions. To address the problem, we retrain and adapt the Teacher by using the same region proposals from Student's detector while the features are from Teacher's own object detector. With aligned network inputs, the adapted Teacher is capable of transferring the knowledge through the intermediate representations. Specifically, we use the mean square error loss to mimic the attention distribution inside the transformer block and present a token-wise noise contrastive loss to align the hidden state by contrasting with negative representations stored in a sample queue. To this end, we show that our proposed distillation significantly improves the performance of small VL models on image captioning and visual question answering tasks. It reaches 120.8 in CIDEr score on COCO captioning, an improvement of 5.1 over its non-distilled counterpart; and an accuracy of 69.8 on VQA 2.0, a 0.8 gain from the baseline. Our extensive experiments and ablations confirm the effective-ness of VL distillation in both pre-training and fine-tuning stages.

We focus on contrastive methods for self-supervised video representation learning. A common paradigm in contrastive learning is to construct positive pairs by sampling different data views for the same instance, with different data instances as negatives. These methods implicitly assume a set of representational invariances to the view selection mechanism (e.g., sampling frames with temporal shifts), which may lead to poor performance on downstream tasks which violate these invariances (fine-grained video action recognition that would benefit from temporal information). To overcome this limitation, we propose an `augmentation aware' contrastive learning framework, where we explicitly provide a sequence of augmentation parameterisations (such as the values of the time shifts used to create data views) as composable augmentation encodings (CATE) to our model when projecting the video representations for contrastive learning. We show that representations learned by our method encode valuable information about specified spatial or temporal augmentation, and in doing so also achieve state-of-the-art performance on a number of video benchmarks.

Weakly Supervised Semantic Segmentation (WSSS) based on image-level labels has been greatly advanced by exploiting the outputs of Class Activation Map (CAM) to generate the pseudo labels for semantic segmentation. However, CAM merely discovers seeds from a small number of regions, which may be insufficient to serve as pseudo masks for semantic segmentation. In this paper, we formulate the expansion of object regions in CAM as an increase in information. From the perspective of information theory, we propose a novel Complementary Patch (CP) Representation and prove that the information of the sum of the CAMs by a pair of input images with complementary hidden (patched) parts, namely CP Pair, is greater than or equal to the information of the baseline CAM. Therefore, a CAM with more information related to object seeds can be obtained by narrowing down the gap between the sum of CAMs generated by the CP Pair and the original CAM. We propose a CP Network (CPN) implemented by a triplet network and three regularization functions. To further improve the quality of the CAMs, we propose a Pixel-Region Correlation Module (PRCM) to augment the contextual information by using object-region relations between the feature maps and the CAMs. Experimental results on the PASCAL VOC 2012 datasets show that our proposed method achieves a new state-of-the-art in WSSS, validating the effectiveness of our CP Representation and CPN.

Traditional approaches for learning 3D object categories have been predominantly trained and evaluated on synthetic datasets due to the unavailability of real 3D-annotated category-centric data. Our main goal is to facilitate advances in this field by collecting real-world data in a magnitude similar to the existing synthetic counterparts. The principal contribution of this work is thus a large-scale dataset, called Common Objects in 3D, with real multi-view images of object categories annotated with camera poses and ground truth 3D point clouds. The dataset contains a total of 1.5 million frames from nearly 19,000 videos capturing objects from 50 MS-COCO categories and, as such, it is significantly larger than alternatives both in terms of the number of categories and objects. We exploit this new dataset to conduct one of the first large-scale "in-the-wild" evaluations of several new-view-synthesis and category-centric 3D reconstruction methods. Finally, we contribute NerFormer - a novel neural rendering method that leverages the powerful Transformer to reconstruct an object given a small number of its views.

We propose a new approach for high resolution semantic image synthesis. It consists of one base image generator and multiple class-specific generators. The base generator generates high quality images based on a segmentation map. To further improve the quality of different objects, we create a bank of Generative Adversarial Networks (GANs) by separately training class-specific models. This has several benefits including -- dedicated weights for each class; centrally aligned data for each model; additional training data from other sources, potential of higher resolution and quality; and easy manipulation of a specific object in the scene. Experiments show that our approach can generate high quality images in high resolution while having flexibility of object-level control by using class-specific generators.

Automatic polyp segmentation from colonoscopy images is an essential step in computer aided diagnosis for colorectal cancer. Most of polyp segmentation methods reported in recent years are based on fully supervised deep learning. However, annotation for polyp images by physicians during the diagnosis is time-consuming and costly. In this paper, we present a novel semi-supervised polyp segmentation via collaborative and adversarial learning of focused and dispersive representations learning model, where focused and dispersive extraction module are used to deal with the diversity of location and shape of polyps. In addition, confidence maps produced by a discriminator in an adversarial training framework shows the effectiveness of leveraging unlabeled data and improving the performance of segmentation network. Consistent regularization is further employed to optimize the segmentation networks to strengthen the representation of the outputs of focused and dispersive extraction module. We also propose an auxiliary adversarial learning method to better leverage unlabeled examples to further improve semantic segmentation accuracy. We conduct extensive experiments on two famous polyp datasets: Kvasir-SEG and CVC-Clinic DB. Experimental results demonstrate the effectiveness of the proposed model, consistently outperforming state-of-the-art semi-supervised segmentation models based on adversarial training and even some advanced fully supervised models. Codes will be released upon publication.

Unsupervised representation learning has achieved outstanding performances using centralized data available on the Internet. However, the increasing awareness of privacy protection limits sharing of decentralized unlabeled image data that grows explosively in multiple parties (e.g. mobile phones and cameras). As such, a natural problem is how to leverage these data to learn visual representations for downstream tasks while preserving data privacy. To address this problem, we propose a novel federated unsupervised learning framework, FedU. In this framework, each party trains models from unlabeled data independently using contrastive learning with an online network and a target network. Then, a central server aggregates trained models and updates clients' models with the aggregated global model. It preserves data privacy as each party only has access to its raw data. Decentralized data among multiple parties is normally non-independent and identically distributed (non-IID), which leads to performance degradation. To tackle this challenge, we propose two simple but effective methods: (1) We design the communication protocol to upload only the encoders of online networks for server aggregation and update them with the aggregated encoder. (2) We introduce a new module to dynamically decide how to update the predictors based on the degree of divergence caused by non-IID. The predictor is the other component of the online network. Extensive experiments and ablations demonstrate the effectiveness and significance of FedU. It outperforms training with only one party by over 5% and other methods by over 14% in linear and semi-supervised evaluation on non-IID data.

Contemporary domain generalization (DG) and multi-source unsupervised domain adaptation (UDA) methods mostly collect data from multiple domains together for joint optimization. However, this centralized training paradigm poses a threat to data privacy and is not applicable when data are non-shared across domains. In this work, we propose a new approach called Collaborative Optimization and Aggregation (COPA), which aims at optimizing a generalized target model for decentralized DG and UDA, where data from different domains are non-shared and private. Our base model consists of a domain-invariant feature extractor and an ensemble of domain-specific classifiers. In an iterative learning process, we optimize a local model for each domain, and then centrally aggregate local feature extractors and assemble domain-specific classifiers to construct a generalized global model, without sharing data from different domains. To improve generalization of feature extractors, we employ hybrid batch-instance normalization and collaboration of frozen classifiers. For better decentralized UDA, we further introduce a prediction agreement mechanism to overcome local disparities towards central model aggregation. Extensive experiments on five DG and UDA benchmark datasets show that COPA is capable of achieving comparable performance against the state-of-the-art DG and UDA methods without the need for centralized data collection in model training.

Typical domain adaptation techniques aim to transfer label information from a label-rich source domain to a label-scarce target domain in the same label space. However, it is often hard to get even the unlabeled target domain data of a task of interest. In such a case, we can capture the domain shift between the source domain and target domain from an unseen task and transfer it to the task of interest, which is known as zero-shot domain adaptation (ZSDA). Existing state-of-the-art methods for ZSDA attempted to generate target domain data. However, training such generative models causes significant computational overhead and is hardly optimized. In this paper, we propose a novel ZSDA method that learns a task-agnostic domain shift by collaborative training of domain-invariant semantic features and task-invariant domain features via adversarial learning. Meanwhile, the spatial attention map is learned from disentangled feature representations to selectively emphasize the domain-specific salient parts of the domain-invariant features. Experimental results show that our ZSDA method achieves state-of-the-art performance on several benchmarks.

CodeNeRF is an implicit 3D neural representation that learns the variation of object shapes and textures across a category and can be trained, from a set of posed images, to synthesize novel views of unseen objects. Unlike the original NeRF, which is scene specific, CodeNeRF learns to disentangle shape and texture by learning separate embeddings. At test time, given a single unposed image of an unseen object, CodeNeRF jointly estimates camera viewpoint, and shape and appearance codes via optimization. Unseen objects can be reconstructed from a single image, and then rendered from new viewpoints or their shape and texture edited by varying the latent codes. We conduct experiments on the SRN benchmark, which show that CodeNeRF generalises well to unseen objects and achieves on-par performance with methods that require known camera pose at test time. Our results on real-world images demonstrate that CodeNeRF can bridge the sim-to-real gap. Project page: https://github.com/wayne1123/code-nerf

Few-shot learning is based on the premise that labels are expensive, especially when they are fine-grained and require expertise. But coarse labels might be easy to acquire and thus abundant. We present a representation learning approach - PAS that allows few-shot learners to leverage coarsely-labeled data available before evaluation. Inspired by self-training, we label the additional data using a teacher trained on the base dataset and filter the teacher's prediction based on the coarse labels; a new student representation is then trained on the base dataset and the pseudo-labeled dataset. PAS is able to produce a representation that consistently and significantly outperforms the baselines in 3 different datasets. Code is available at https://github.com/cpphoo/PAS.

Semi-supervised learning has been an effective paradigm for leveraging unlabeled data to reduce the reliance on labeled data. We propose CoMatch, a new semi-supervised learning method that unifies dominant approaches and addresses their limitations. CoMatch jointly learns two representations of the training data, their class probabilities and low-dimensional embeddings. The two representations interact with each other to jointly evolve. The embeddings impose a smoothness constraint on the class probabilities to improve the pseudo-labels, whereas the pseudo-labels regularize the structure of the embeddings through graph-based contrastive learning. CoMatch achieves state-of-the-art performance on multiple datasets. It achieves substantial accuracy improvements on the label-scarce CIFAR-10 and STL-10. On ImageNet with 1% labels, CoMatch achieves a top-1 accuracy of 66.0%, outperforming FixMatch by 12.6%. Furthermore, CoMatch achieves better representation learning performance on downstream tasks, outperforming both supervised learning and self-supervised learning. Code and pre-trained models are available at https://github.com/salesforce/CoMatch/.

Recent breakthroughs in self-supervised learning show that such algorithms learn visual representations that can be transferred better to unseen tasks than cross-entropy based methods which rely on task-specific supervision. In this paper, we found that the similar holds in the continual learning context: contrastively learned representations are more robust against the catastrophic forgetting than ones trained with the cross-entropy objective. Based on this novel observation, we propose a rehearsal-based continual learning algorithm that focuses on continually learning and maintaining transferable representations. More specifically, the proposed scheme (1) learns representations using the contrastive learning objective, and (2) preserves learned representations using a self-supervised distillation step. We conduct extensive experimental validations under popular benchmark image classification datasets, where our method sets the new state-of-the-art performance. Source code is available at https://github.com/chaht01/Co2L.