CVPR 2022 论文列表

Multi-modal learning from video data has seen increased attention recently as it allows training of semantically meaningful embeddings without human annotation, enabling tasks like zero-shot retrieval and action localization. In this work, we present a multi-modal, modality agnostic fusion transformer that learns to exchange information between multiple modalities, such as video, audio, and text, and integrate them into a fused representation in a joined multi-modal embedding space. We propose to train the system with a combinatorial loss on everything at once - any combination of input modalities, such as single modalities as well as pairs of modalities, explicitly leaving out any add-ons such as position or modality encoding. At test time, the resulting model can process and fuse any number of input modalities. Moreover, the implicit properties of the transformer allow to process inputs of different lengths. To evaluate the proposed approach, we train the model on the large scale HowTo100M dataset and evaluate the resulting embedding space on four challenging benchmark datasets obtaining state-of-the-art results in zero-shot video retrieval and zero-shot video action localization. Our code for this work is also available.

Neuromorphic vision sensor is a new bio-inspired imaging paradigm that reports asynchronous, continuously per-pixel brightness changes called 'events' with high temporal resolution and high dynamic range. So far, the event-based image reconstruction methods are based on artificial neural networks (ANN) or hand-crafted spatiotemporal smoothing techniques. In this paper, we first implement the image reconstruction work via deep spiking neural network (SNN) architecture. As the bio-inspired neural networks, SNNs operating with asynchronous binary spikes distributed over time, can potentially lead to greater computational efficiency on event-driven hardware. We propose a novel Event-based Video reconstruction framework based on a fully Spiking Neural Network (EVSNN), which utilizes Leaky-Integrate-and-Fire (LIF) neuron and Membrane Potential (MP) neuron. We find that the spiking neurons have the potential to store useful temporal information (memory) to complete such time-dependent tasks. Furthermore, to better utilize the temporal information, we propose a hybrid potential-assisted framework (PA-EVSNN) using the membrane potential of spiking neuron. The proposed neuron is referred as Adaptive Membrane Potential (AMP) neuron, which adaptively updates the membrane potential according to the input spikes. The experimental results demonstrate that our models achieve comparable performance to ANN-based models on IJRR, MVSEC, and HQF datasets. The energy consumptions of EVSNN and PA-EVSNN are 19.36 times and 7.75 times more computationally efficient than their ANN architectures, respectively. The code and pretrained model are available at https://sites.google.com/view/evsnn.

We introduce EDS, a direct monocular visual odometry using events and frames. Our algorithm leverages the event generation model to track the camera motion in the blind time between frames. The method formulates a direct probabilistic approach of observed brightness increments. Per-pixel brightness increments are predicted using a sparse number of selected 3D points and are compared to the events via the brightness increment error to estimate camera motion. The method recovers a semi-dense 3D map using photometric bundle adjustment. EDS is the first method to perform 6-DOF VO using events and frames with a direct approach. By design it overcomes the problem of changing appearance in indirect methods. Our results outperform all previous event-based odometry solutions. We also show that, for a target error performance, EDS can work at lower frame rates than state-of-the-art frame-based VO solutions. This opens the door to low-power motion-tracking applications where frames are sparingly triggered "on demand" and our method tracks the motion in between. We release code and datasets to the public.

Knowledge distillation (KD) is a widely-used technique that utilizes large networks to improve the performance of compact models. Previous KD approaches usually aim to guide the student to mimic the teacher's behavior completely in the representation space. However, such one-to-one corresponding constraints may lead to inflexible knowledge transfer from the teacher to the student, especially those with low model capacities. Inspired by the ultimate goal of KD methods, we propose a novel Evaluation oriented KD method (EKD) for deep face recognition to directly reduce the performance gap between the teacher and student models during training. Specifically, we adopt the commonly used evaluation metrics in face recognition, i.e., False Positive Rate (FPR) and True Positive Rate (TPR) as the performance indicator. According to the evaluation protocol, the critical pair relations that cause the TPR and FPR difference between the teacher and student models are selected. Then, the critical relations in the student are constrained to approximate the corresponding ones in the teacher by a novel rank-based loss function, giving more flexibility to the student with low capacity. Extensive experimental results on popular benchmarks demonstrate the superiority of our EKD over state-of-the-art competitors.

Neural networks trained with SGD were recently shown to rely preferentially on linearly-predictive features and can ignore complex, equally-predictive ones. This simplicity bias can explain their lack of robustness out of distribution (OOD). The more complex the task to learn, the more likely it is that statistical artifacts (i.e. selection biases, spurious correlations) are simpler than the mechanisms to learn. We demonstrate that the simplicity bias can be mitigated and OOD generalization improved. We train a set of similar models to fit the data in different ways using a penalty on the alignment of their input gradients. We show theoretically and empirically that this induces the learning of more complex predictive patterns. OOD generalization fundamentally requires information beyond i.i.d. examples, such as multiple training environments, counterfactual examples, or other side information. Our approach shows that we can defer this requirement to an independent model selection stage. We obtain SOTA results in visual recognition on biased data and generalization across visual domains. The method - the first to evade the simplicity bias - highlights the need for a better understanding and control of inductive biases in deep learning.

This paper proposes to use neuromorphic events for correcting rolling shutter (RS) images as consecutive global shutter (GS) frames. RS effect introduces edge distortion and region occlusion into images caused by row-wise readout of CMOS sensors. We introduce a novel computational imaging setup consisting of an RS sensor and an event sensor, and propose a neural network called EvUnroll to solve this problem by exploring the high-temporal-resolution property of events. We use events to bridge a spatio-temporal connection between RS and GS, establish a flow estimation module to correct edge distortions, and design a synthesis-based restoration module to restore occluded regions. The results of two branches are fused through a refining module to generate corrected GS images. We further propose datasets captured by a high-speed camera and an RS-Event hybrid camera system for training and testing our network. Experimental results on both public and proposed datasets show a systematic performance improvement compared to state-of-the-art methods.

We introduce Ev-TTA, a simple, effective test-time adaptation algorithm for event-based object recognition. While event cameras are proposed to provide measurements of scenes with fast motions or drastic illumination changes, many existing event-based recognition algorithms suffer from performance deterioration under extreme conditions due to significant domain shifts. Ev-TTA mitigates the severe domain gaps by fine-tuning the pre-trained classifiers during the test phase using loss functions inspired by the spatio-temporal characteristics of events. Since the event data is a temporal stream of measurements, our loss function enforces similar predictions for adjacent events to quickly adapt to the changed environment online. Also, we utilize the spatial correlations between two polarities of events to handle noise under extreme illumination, where different polarities of events exhibit distinctive noise distributions. Ev-TTA demonstrates a large amount of performance gain on a wide range of event-based object recognition tasks without extensive additional training. Our formulation can be successfully applied regardless of input representations and further extended into regression tasks. We expect Ev-TTA to provide the key technique to deploy event-based vision algorithms in challenging real-world applications where significant domain shift is inevitable.

In machine learning, disparity metrics are often defined by measuring the difference in the performance or outcome of a model, across different sub-populations (groups) of datapoints. Thus, the inputs to disparity quantification consist of a model's predictions y', the ground-truth labels for the predictions y, and group labels g for the data points. Performance of the model for each group is calculated by comparing y' and y for the datapoints within a specific group, and as a result, disparity of performance across the different groups can be calculated. In many real world scenarios however, group labels (g) may not be available at scale during training and validation time, or collecting them might not be feasible or desirable as they could often be sensitive information. As a result, evaluating disparity metrics across categorical groups would not be feasible. On the other hand, in many scenarios noisy groupings may be obtainable using some form of a proxy, which would allow measuring disparity metrics across sub-populations. Here we explore performing such analysis on computer vision models trained on human faces, and on tasks such as face attribute prediction and affect estimation. Our experiments indicate that embeddings resulting from an off-the-shelf face recognition model, could meaningfully serve as a proxy for such estimation.

Image denoising has achieved unprecedented progress as great efforts have been made to exploit effective deep denoisers. To improve the denoising performance in real-world, two typical solutions are used in recent trends: devising better noise models for the synthesis of more realistic training data, and estimating noise level function to guide non-blind denoisers. In this work, we combine both noise modeling and estimation, and propose an innovative noise model estimation and noise synthesis pipeline for realistic noisy image generation. Specifically, our model learns a noise estimation model with fine-grained statistical noise model in a contrastive manner. Then, we use the estimated noise parameters to model camera-specific noise distribution, and synthesize realistic noisy training data. The most striking thing for our work is that by calibrating noise models of several sensors, our model can be extended to predict other cameras. In other words, we can estimate camera-specific noise models for unknown sensors with only testing images, without any laborious calibration frames or paired noisy/clean data. The proposed pipeline endows deep denoisers with competitive performances with state-of-the-art real noise modeling methods.

In machine learning, a question of great interest is understanding what examples are challenging for a model to classify. Identifying atypical examples ensures the safe deployment of models, isolates samples that require further human inspection, and provides interpretability into model behavior. In this work, we propose Variance of Gradients (VoG) as a valuable and efficient metric to rank data by difficulty and to surface a tractable subset of the most challenging examples for human-in-the-loop auditing. We show that data points with high VoG scores are far more difficult for the model to learn and over-index on corrupted or memorized examples. Further, restricting the evaluation to the test set instances with the lowest VoG improves the model's generalization performance. Finally, we show that VoG is a valuable and efficient ranking for out-of-distribution detection

Egocentric 3D human pose estimation with a single fisheye camera has drawn a significant amount of attention recently. However, existing methods struggle with pose estimation from in-the-wild images, because they can only be trained on synthetic data due to the unavailability of large-scale in-the-wild egocentric datasets. Furthermore, these methods easily fail when the body parts are occluded by or interacting with the surrounding scene. To address the shortage of in-the-wild data, we collect a large-scale in-the-wild egocentric dataset called Egocentric Poses in the Wild (EgoPW). This dataset is captured by a head-mounted fisheye camera and an auxiliary external camera, which provides an additional observation of the human body from a third-person perspective during training. We present a new egocentric pose estimation method, which can be trained on the new dataset with weak external supervision. Specifically, we first generate pseudo labels for the EgoPW dataset with a spatio-temporal optimization method by incorporating the external-view supervision. The pseudo labels are then used to train an egocentric pose estimation network. To facilitate the network training, we propose a novel learning strategy to supervise the egocentric features with the high-quality features extracted by a pretrained external-view pose estimation model. The experiments show that our method predicts accurate 3D poses from a single in-the-wild egocentric image and outperforms the state-of-the-art methods both quantitatively and qualitatively.

In Heterogeneous Face Recognition (HFR), the objective is to match faces across two different domains such as visible and thermal. Large domain discrepancy makes HFR a difficult problem. Recent methods attempting to fill the gap via synthesis have achieved promising results, but their performance is still limited by the scarcity of paired training data. In practice, large-scale heterogeneous face data are often inaccessible due to the high cost of acquisition and annotation process as well as privacy regulations. In this paper, we propose a new face hallucination paradigm for HFR, which not only enables data-efficient synthesis but also allows to scale up model training without breaking any privacy policy. Unlike existing methods that learn face synthesis entirely from scratch, our approach is particularly designed to take advantage of rich and diverse facial priors from visible domain for more faithful hallucination. On the other hand, large-scale training is enabled by introducing a new federated learning scheme to allow institution-wise collaborations while avoiding explicit data sharing. Extensive experiments demonstrate the advantages of our approach in tackling HFR under current data limitations. In a unified framework, our method yields the state-of-the-art hallucination results on multiple HFR datasets.

Equivariance has been a long-standing concern in various fields ranging from computer vision to physical modeling. Most previous methods struggle with generality, simplicity, and expressiveness --- some are designed ad hoc for specific data types, some are too complex to be accessible, and some sacrifice flexible transformations. In this work, we propose a novel and simple framework to achieve equivariance for point cloud analysis based on the message passing (graph neural network) scheme. We find the equivariant property could be obtained by introducing an orientation for each point to decouple the relative position for each point from the global pose of the entire point cloud. Therefore, we extend current message passing networks with a module that learns orientations for each point. Before aggregating information from the neighbors of a point, the networks transforms the neighbors' coordinates based on the point's learned orientations. We provide formal proofs to show the equivariance of the proposed framework. Empirically, we demonstrate that our proposed method is competitive on both point cloud analysis and physical modeling tasks. Code is available at https://github.com/luost26/Equivariant-OrientedMP.

Pooling multiple neuroimaging datasets across institutions often enables significant improvements in statistical power when evaluating associations (e.g., between risk factors and disease outcomes) that would otherwise be too weak to detect. When there is only a single source of variability (e.g., different scanners), domain adaptation and matching the distributions of representations may suffice in many scenarios. But in the presence of more than one nuisance variable which concurrently influence the measurements, pooling datasets poses unique challenges, e.g., variations in the data can come from both the acquisition method as well as the demographics of participants (gender, age). Invariant representation learning, by itself, is ill-suited to fully model the data generation process. In this paper, we show how bringing recent results on equivariant representation learning (for studying symmetries in neural networks) together with simple use of classical results on causal inference provides an effective practical solution to this problem. In particular, we demonstrate how our model allows dealing with more than one nuisance variable under some assumptions and can enable (relatively) painless analysis of pooled scientific datasets in scenarios that would otherwise entail removing a large portion of the samples.

Despite the recent success of long-tailed object detection, almost all long-tailed object detectors are developed based on the two-stage paradigm. In practice, one-stage detectors are more prevalent in the industry because they have a simple and fast pipeline that is easy to deploy. However, in the long-tailed scenario, this line of work has not been explored so far. In this paper, we investigate whether one-stage detectors can perform well in this case. We discover the primary obstacle that prevents one-stage detectors from achieving excellent performance is: categories suffer from different degrees of positive-negative imbalance problems under the long-tailed data distribution. The conventional focal loss balances the training process with the same modulating factor for all categories, thus failing to handle the long-tailed problem. To address this issue, we propose the Equalized Focal Loss (EFL) that rebalances the loss contribution of positive and negative samples of different categories independently according to their imbalance degrees. Specifically, EFL adopts a category-relevant modulating factor which can be adjusted dynamically by the training status of different categories. Extensive experiments conducted on the challenging LVIS v1 benchmark demonstrate the effectiveness of our proposed method. With an end-to-end training pipeline, EFL achieves 29.2% in terms of overall AP and obtains significant performance improvements on rare categories, surpassing all existing state-of-the-art methods. The code is available at https://github.com/ModelTC/EOD.

Egocentric augmented reality devices such as wearable glasses passively capture visual data as a human wearer tours a home environment. We envision a scenario wherein the human communicates with an AI agent powering such a device by asking questions (e.g., "where did you last see my keys?"). In order to succeed at this task, the egocentric AI assistant must (1) construct semantically rich and efficient scene memories that encode spatio-temporal information about objects seen during the tour and (2) possess the ability to understand the question and ground its answer into the semantic memory representation. Towards that end, we introduce (1) a new task -- Episodic Memory Question Answering (EMQA) wherein an egocentric AI assistant is provided with a video sequence (the tour) and a question as an input and is asked to localize its answer to the question within the tour, (2) a dataset of grounded questions designed to probe the agent's spatio-temporal understanding of the tour, and (3) a model for the task that encodes the scene as an allocentric, top-down semantic feature map and grounds the question into the map to localize the answer. We show that our choice of episodic scene memory outperforms naive, off-the-shelf solutions for the task as well as a host of very competitive baselines and is robust to noise in depth, pose as well as camera jitter.

In Vision-and-Language Navigation (VLN), an agent needs to navigate through the environment based on natural language instructions. Due to limited available data for agent training and finite diversity in navigation environments, it is challenging for the agent to generalize to new, unseen environments. To address this problem, we propose EnvEdit, a data augmentation method that creates new environments by editing existing environments, which are used to train a more generalizable agent. Our augmented environments can differ from the seen environments in three diverse aspects: style, object appearance, and object classes. Training on these edit-augmented environments prevents the agent from overfitting to existing environments and helps generalize better to new, unseen environments. Empirically, on both the Room-to-Room and the multi-lingual Room-Across-Room datasets, we show that our proposed EnvEdit method gets significant improvements in all metrics on both pre-trained and non-pre-trained VLN agents, and achieves the new state-of-the-art on the test leaderboard. We further ensemble the VLN agents augmented on different edited environments and show that these edit methods are complementary.

Active learning is a promising alternative to alleviate the issue of high annotation cost in the computer vision tasks by consciously selecting more informative samples to label. Active learning for object detection is more challenging and existing efforts on it are relatively rare. In this paper, we propose a novel hybrid approach to address this problem, where the instance-level uncertainty and diversity are jointly considered in a bottom-up manner. To balance the computational complexity, the proposed approach is designed as a two-stage procedure. At the first stage, an Entropy-based Non-Maximum Suppression (ENMS) is presented to estimate the uncertainty of every image, which performs NMS according to the entropy in the feature space to remove predictions with redundant information gains. At the second stage, a diverse prototype (DivProto) strategy is explored to ensure the diversity across images by progressively converting it into the intra-class and inter-class diversities of the entropy-based class-specific prototypes. Extensive experiments are conducted on MS COCO and Pascal VOC, and the proposed approach achieves state of the art results and significantly outperforms the other counterparts, highlighting its superiority.

The advent of large-scale training has produced a cornucopia of powerful visual recognition models. However, generative models, such as GANs, have traditionally been trained from scratch in an unsupervised manner. Can the collective "knowledge" from a large bank of pretrained vision models be leveraged to improve GAN training? If so, with so many models to choose from, which one(s) should be selected, and in what manner are they most effective? We find that pretrained computer vision models can significantly improve performance when used in an ensemble of discriminators. Notably, the particular subset of selected models greatly affects performance. We propose an effective selection mechanism, by probing the linear separability between real and fake samples in pretrained model embeddings, choosing the most accurate model, and progressively adding it to the discriminator ensemble. Interestingly, our method can improve GAN training in both limited data and large-scale settings. Given only 10k training samples, our FID on LSUN Cat matches the StyleGAN2 trained on 1.6M images. On the full dataset, our method improves FID by 1.5 to 2 times on cat, church, and horse categories of LSUN.