CVPR 2022 Papers

Appearance-based Gaze Estimation leverages deep neural networks to regress the gaze direction from monocular images and achieve impressive performance. However, its success depends on expensive and cumbersome annotation capture. When lacking precise annotation, the large domain gap hinders the performance of trained models on new domains. In this paper, we propose a novel gaze adaptation approach, namely Contrastive Regression Gaze Adaptation (CRGA), for generalizing gaze estimation on the target domain in an unsupervised manner. CRGA leverages the Contrastive Domain Generalization (CDG) module to learn the stable representation from the source domain and leverages the Contrastive Self-training Adaptation (CSA) module to learn from the pseudo labels on the target domain. The core of both CDG and CSA is the Contrastive Regression (CR) loss, a novel contrastive loss for regression by pulling features with closer gaze directions closer together while pushing features with farther gaze directions farther apart. Experimentally, we choose ETH-XGAZE and Gaze-360 as the source domain and test the domain generalization and adaptation performance on MPIIGAZE, RT-GENE, GazeCapture, EyeDiap respectively. The results demonstrate that our CRGA achieves remarkable performance improvement compared with the baseline models and also outperforms the state-of-the-art domain adaptation approaches on gaze adaptation tasks.

Video highlight detection can greatly simplify video browsing, potentially paving the way for a wide range of applications. Existing efforts are mostly fully-supervised, requiring humans to manually identify and label the interesting moments (called highlights) in a video. Recent weakly supervised methods forgo the use of highlight annotations, but typically require extensive efforts in collecting external data such as web-crawled videos for model learning. This observation has inspired us to consider unsupervised highlight detection where neither frame-level nor video-level annotations are available in training. We propose a simple contrastive learning framework for unsupervised highlight detection. Our framework encodes a video into a vector representation by learning to pick video clips that help to distinguish it from other videos via a contrastive objective using dropout noise. This inherently allows our framework to identify video clips corresponding to highlight of the video. Extensive empirical evaluations on three highlight detection benchmarks demonstrate the superior performance of our approach.

We propose a novel probabilistic method employing Bayesian Model Averaging and self-cycle regularization for spatio-temporal correspondence learning in videos within a self-supervised learning framework. Most existing methods for self-supervised correspondence learning suffer from noisy labels that come with the data for free, and the presence of occlusion exacerbates the problem. We tackle this issue within a probabilistic framework that handles model uncertainty inherent in the path selection problem built on a complete graph. We propose a self-cycle regularization to consider a cycle-consistency property on individual edges in order to prevent converging on noisy matching or trivial solutions. We also utilize a mixture of sequential Bayesian filters to estimate posterior distribution for targets. In addition, we present a domain contrastive loss to learn discriminative representation among videos. Our algorithm is evaluated on various datasets for video label propagation tasks including DAVIS2017, VIP and JHMDB, and shows outstanding performances compared to the state-of-the-art self-supervised learning based video correspondence algorithms. Moreover, our method converges significantly faster than previous methods.

Contrastive learning (or its variants) has recently become a promising direction in the self-supervised learning domain, achieving similar performance as supervised learning with minimum fine-tuning. Despite the labeling efficiency, wide and large networks are required to achieve high accuracy, which incurs a high amount of computation and hinders the pragmatic merit of self-supervised learning. To effectively reduce the computation of insignificant features or channels, recent dynamic pruning algorithms for supervised learning employed auxiliary salience predictors. However, we found that such salience predictors cannot be easily trained when they are naively applied to contrastive learning from scratch. To address this issue, we propose contrastive dual gating(CDG), a novel dynamic pruning algorithm that skips the uninformative features during contrastive learning without hurting the trainability of the networks. We demonstrate the superiority of CDG with ResNet models for CIFAR-10, CIFAR-100, and ImageNet-100 datasets. Compared to our implementations of state-of-the-art dynamic pruning algorithms for self-supervised learning, CDG achieves up to 15% accuracy improvement for CIFAR-10 dataset with higher computation reduction.

Conditional Neural Processes(CNPs) bridge neural networks with probabilistic inference to approximate functions of Stochastic Processes under meta-learning settings. Given a batch of non-i.i.d function instantiations, CNPs are jointly optimized for in-instantiation observation prediction and cross-instantiation meta-representation adaptation within a generative reconstruction pipeline. There can be a challenge in tying together such two targets when the distribution of function observations scales to high-dimensional and noisy spaces. Instead, noise contrastive estimation might be able to provide more robust representations by learning distributional matching objectives to combat such inherent limitation of generative models. In light of this, we propose to equip CNPs by 1) aligning prediction with encoded ground-truth observation, and 2) decoupling meta-representation adaptation from generative reconstruction. Specifically, two auxiliary contrastive branches are set up hierarchically, namely in-instantiation temporal contrastive learning (TCL) and cross-instantiation function contrastive learning (FCL), to facilitate local predictive alignment and global function consistency, respectively. We empirically show that TCL captures high-level abstraction of observations, whereas FCL helps identify underlying functions, which in turn provides more efficient representations. Our model outperforms other CNPs variants when evaluating function distribution reconstruction and parameter identification across 1D, 2D and high-dimensional time-series.

Point cloud segmentation is fundamental in understanding 3D environments. However, current 3D point cloud segmentation methods usually perform poorly on scene boundaries, which degenerates the overall segmentation performance. In this paper, we focus on the segmentation of scene boundaries. Accordingly, we first explore metrics to evaluate the segmentation performance on scene boundaries. To address the unsatisfactory performance on boundaries, we then propose a novel contrastive boundary learning (CBL) framework for point cloud segmentation. Specifically, the proposed CBL enhances feature discrimination between points across boundaries by contrasting their representations with the assistance of scene contexts at multiple scales. By applying CBL on three different baseline methods, we experimentally show that CBL consistently improves different baselines and assists them to achieve compelling performance on boundaries, as well as the overall performance, e.g. in mIoU. The experimental results demonstrate the effectiveness of our method and the importance of boundaries for 3D point cloud segmentation. Code and model will be made publicly available at https://github.com/LiyaoTang/contrastBoundary.

Partially-supervised instance segmentation is a task which requests segmenting objects from novel categories via learning on limited base categories with annotated masks thus eliminating demands of heavy annotation burden. The key to addressing this task is to build an effective class-agnostic mask segmentation model. Unlike previous methods that learn such models only on base categories, in this paper, we propose a new method, named ContrastMask, which learns a mask segmentation model on both base and novel categories under a unified pixel-level contrastive learning framework. In this framework, annotated masks of base categories and pseudo masks of novel categories serve as a prior for contrastive learning, where features from the mask regions (foreground) are pulled together, and are contrasted against those from the background, and vice versa. Through this framework, feature discrimination between foreground and background is largely improved, facilitating learning of the class-agnostic mask segmentation model. Exhaustive experiments on the COCO dataset demonstrate the superiority of our method, which outperforms previous state-of-the-arts.

We propose an effective and easy-to-implement method for simultaneously performing landmark detection in images and obtaining an ingenious uncertainty measurement for each landmark. Uncertainty measurements for landmarks are particularly useful in medical imaging applications: rather than giving an erroneous reading, a landmark detection system is more useful when it flags its level of confidence in its prediction. When an automated system is unsure of its predictions, the accuracy of the results can be further improved manually by a human. In the medical domain, being able to review an automated system's level of certainty significantly improves a clinician's trust in it. This paper obtains landmark predictions with uncertainty measurements using a three stage method: 1) We train our network on one-hot heatmap images, 2) We calibrate the uncertainty of the network using temperature scaling, 3) We calculate a novel statistic called 'Expected Radial Error' to obtain uncertainty measurements. We find that this method not only achieves localisation results on par with other state-of-the-art methods but also an uncertainty score which correlates with the true error for each landmark thereby bringing an overall step change in what a generic computer vision method for landmark detection should be capable of. In addition, we show that our uncertainty measurement can be used to classify, with good accuracy, what landmark predictions are likely to be inaccurate. Code available at: https://github.com/jfm15/ContourHuggingHeatmaps.git

We propose Continuous Scene Representations (CSR), a scene representation constructed by an embodied agent navigating within a space, where objects and their relationships are modeled by continuous valued embeddings. Our method captures feature relationships between objects, composes them into a graph structure on-the-fly, and situates an embodied agent within the representation. Our key insight is to embed pair-wise relationships between objects in a latent space. This allows for a richer representation compared to discrete relations (e.g., [support], [next-to]) commonly used for building scene representations. CSR can track objects as the agent moves in a scene, update the representation accordingly, and detect changes in room configurations. Using CSR, we outperform state-of-the-art approaches for the challenging downstream task of visual room rearrangement, without any task specific training. Moreover, we show the learned embeddings capture salient spatial details of the scene and show applicability to real world data. A summery video and code is available at https://prior.allenai.org/projects/csr.

Test-time domain adaptation aims to adapt a source pre-trained model to a target domain without using any source data. Existing works mainly consider the case where the target domain is static. However, real-world machine perception systems are running in non-stationary and continually changing environments where the target domain distribution can change over time. Existing methods, which are mostly based on self-training and entropy regularization, can suffer from these non-stationary environments. Due to the distribution shift over time in the target domain, pseudo-labels become unreliable. The noisy pseudo-labels can further lead to error accumulation and catastrophic forgetting. To tackle these issues, we propose a continual test-time adaptation approach (CoTTA) which comprises two parts. Firstly, we propose to reduce the error accumulation by using weight-averaged and augmentation-averaged predictions which are often more accurate. On the other hand, to avoid catastrophic forgetting, we propose to stochastically restore a small part of the neurons to the source pre-trained weights during each iteration to help preserve source knowledge in the long-term. The proposed method enables the long-term adaptation for all parameters in the network. CoTTA is easy to implement and can be readily incorporated in off-the-shelf pre-trained models. We demonstrate the effectiveness of our approach on four classification tasks and a segmentation task for continual test-time adaptation, on which we outperform existing methods. Our code is available at https://qin.ee/cotta.

Predictive learning ideally builds the world model of physical processes in one or more given environments. Typical setups assume that we can collect data from all environments at all times. In practice, however, different prediction tasks may arrive sequentially so that the environments may change persistently throughout the training procedure. Can we develop predictive learning algorithms that can deal with more realistic, non-stationary physical environments? In this paper, we study a new continual learning problem in the context of video prediction, and observe that most existing methods suffer from severe catastrophic forgetting in this setup. To tackle this problem, we propose the continual predictive learning (CPL) approach, which learns a mixture world model via predictive experience replay and performs test-time adaptation with non-parametric task inference. We construct two new benchmarks based on RoboNet and KTH, in which different tasks correspond to different physical robotic environments or human actions. Our approach is shown to effectively mitigate forgetting and remarkably outperform the naive combinations of previous art in video prediction and continual learning.

Continual learning is a challenging real-world problem for constructing a mature AI system when data are provided in a streaming fashion. Despite recent progress in continual classification, the researches of continual object detection are impeded by the diverse sizes and numbers of objects in each image. Different from previous works that tune the whole network for all tasks, in this work, we present a simple and flexible framework for continual object detection via pRotOtypical taSk corrElaTion guided gaTing mechAnism (ROSETTA). Concretely, a unified framework is shared by all tasks while task-aware gates are introduced to automatically select sub-models for specific tasks. In this way, various knowledge can be successively memorized by storing their corresponding sub-model weights in this system. To make ROSETTA automatically determine which experience is available and useful, a prototypical task correlation guided Gating Diversity Controller (GDC) is introduced to adaptively adjust the diversity of gates for the new task based on class-specific prototypes. GDC module computes class-to-class correlation matrix to depict the cross-task correlation, and hereby activates more exclusive gates for the new task if a significant domain gap is observed. Comprehensive experiments on COCO-VOC, KITTI-Kitchen, class-incremental detection on VOC and sequential learning of four tasks show that ROSETTA yields state-of-the-art performance on both task-based and class-based continual object detection.

In visual search, the gallery set could be incrementally growing and added to the database in practice. However, existing methods rely on the model trained on the entire dataset, ignoring the continual updating of the model. Besides, as the model updates, the new model must re-extract features for the entire gallery set to maintain compatible feature space, imposing a high computational cost for a large gallery set. To address the issues of long-term visual search, we introduce a continual learning (CL) approach that can handle the incrementally growing gallery set with backward embedding consistency. We enforce the losses of inter-session data coherence, neighbor-session model coherence, and intra-session discrimination to conduct a continual learner. In addition to the disjoint setup, our CL solution also tackles the situation of increasingly adding new classes for the blurry boundary without assuming all categories known in the beginning and during model update. To our knowledge, this is the first CL method both tackling the issue of backward-consistent feature embedding and allowing novel classes to occur in the new sessions. Extensive experiments on various benchmarks show the efficacy of our approach under a wide range of setups.

Continual learning methods aim at training a neural network from sequential data with streaming labels, relieving catastrophic forgetting. However, existing methods are based on and designed for convolutional neural networks (CNNs), which have not utilized the full potential of newly emerged powerful vision transformers. In this paper, we propose a novel attention-based framework Lifelong Vision Transformer (LVT), to achieve a better stability-plasticity trade-off for continual learning. Specifically, an inter-task attention mechanism is presented in LVT, which implicitly absorbs the previous tasks' information and slows down the drift of important attention between previous tasks and the current task. LVT designs a dual-classifier structure that independently injects new representation to avoid catastrophic interference and accumulates the new and previous knowledge in a balanced manner to improve the overall performance. Moreover, we develop a confidence-aware memory update strategy to deepen the impression of the previous tasks. The extensive experimental results show that our approach achieves state-of-the-art performance with even fewer parameters on continual learning benchmarks.

Modern self-supervised learning algorithms typically enforce persistency of instance representations across views. While being very effective on learning holistic image and video representations, such an objective becomes suboptimal for learning spatio-temporally fine-grained features in videos, where scenes and instances evolve through space and time. In this paper, we present Contextualized Spatio-Temporal Contrastive Learning (ConST-CL) to effectively learn spatio-temporally fine-grained video representations via self-supervision. We first design a region-based pretext task which requires the model to transform instance representations from one view to another, guided by context features. Further, we introduce a simple network design that successfully reconciles the simultaneous learning process of both holistic and local representations. We evaluate our learned representations on a variety of downstream tasks and show that ConST-CL achieves competitive results on 6 datasets, including Kinetics, UCF, HMDB, AVAKinetics, AVA and OTB. Our code and models will be available.

Asymmetric image retrieval, which typically uses small model for query side and large model for database server, is an effective solution for resource-constrained scenarios. However, existing approaches either fail to achieve feature coherence or make strong assumptions, e.g., requiring labeled datasets or classifiers from large model, etc., which limits their practical application. To this end, we propose a flexible contextual similarity distillation framework to enhance the small query model and keep its output feature compatible with that of large gallery model, which is crucial with asymmetric retrieval. In our approach, we learn the small model with a new contextual similarity consistency constraint without any data label. During the small model learning, it preserves the contextual similarity among each training image and its neighbors with the features extracted by the large model. Note that this simple constraint is consistent with simultaneous first-order feature vector preserving and second-order ranking list preserving. Extensive experiments show that the proposed method outperforms the state-of-the-art methods on the Revisited Oxford and Paris datasets.

We study the problem of contextual outpainting, which aims to hallucinate the missing background contents based on the remaining foreground contents. Existing image outpainting methods focus on completing object shapes or extending existing scenery textures, neglecting the semantically meaningful relationship between the missing and remaining contents. To explore the semantic cues provided by the remaining foreground contents, we propose a novel ConTextual Outpainting GAN (CTO-GAN), leveraging the semantic layout as a bridge to synthesize coherent and diverse background contents. To model the contextual correlation between foreground and background contents, we incorporate an object-level contrastive loss to regularize the learning of cross-modal representations of foreground contents and the corresponding background semantic layout, facilitating accurate semantic reasoning. Furthermore, we improve the realism of the generated background contents via detecting generated context in adversarial training. Extensive experiments demonstrate that the proposed method achieves superior performance compared with existing solutions on the challenging COCO-stuff dataset.

Crowded scenes make it challenging to differentiate persons and locate their pose keypoints. This paper proposes the Contextual Instance Decoupling (CID), which presents a new pipeline for multi-person pose estimation. Instead of relying on person bounding boxes to spatially differentiate persons, CID decouples persons in an image into multiple instance-aware feature maps. Each of those feature maps is hence adopted to infer keypoints for a specific person. Compared with bounding box detection, CID is differentiable and robust to detection errors. Decoupling persons into different feature maps allows to isolate distractions from other persons, and explore context cues at scales larger than the bounding box size. Experiments show that CID outperforms previous multi-person pose estimation pipelines on crowded scenes pose estimation benchmarks in both accuracy and efficiency. For instance, it achieves 71.3% AP on CrowdPose, outperforming the recent single-stage DEKR by 5.6%, the bottom-up CenterAttention by 3.7%, and the top-down JC-SPPE by 5.3%. This advantage sustains on the commonly used COCO benchmark.