CVPR 2022 Papers

We illustrate the detrimental effect, such as overconfident decisions, that exponential behavior can have in methods like classical LDA and logistic regression. We then show how polynomiality can remedy the situation. This, among others, leads purposefully to random-level performance in the tails, away from the bulk of the training data. A directly related, simple, yet important technical novelty we subsequently present is softRmax: a reasoned alternative to the standard softmax function employed in contemporary (deep) neural networks. It is derived through linking the standard softmax to Gaussian class-conditional models, as employed in LDA, and replacing those by a polynomial alternative. We show that two aspects of softRmax, conservativeness and inherent gradient regularization, lead to robustness against adversarial attacks without gradient obfuscation.

Adversarial training has been shown to be one of the most effective approaches to improve the robustness of deep neural networks. It is formalized as a min-max optimization over model weights and adversarial perturbations, where the weights can be optimized through gradient descent methods like SGD. In this paper, we show that treating model weights as random variables allows for enhancing adversarial training through Second-Order Statistics Optimization (S^2O) with respect to the weights. By relaxing a common (but unrealistic) assumption of previous PAC-Bayesian frameworks that all weights are statistically independent, we derive an improved PAC-Bayesian adversarial generalization bound, which suggests that optimizing second-order statistics of weights can effectively tighten the bound. In addition to this theoretical insight, we conduct an extensive set of experiments, which show that S^2O not only improves the robustness and generalization of the trained neural networks when used in isolation, but also integrates easily in state-of-the-art adversarial training techniques like TRADES, AWP, MART, and AVMixup, leading to a measurable improvement of these techniques. The code is available at https://github.com/Alexkael/S2O.

Owing to security implications of adversarial vulnerability, adversarial robustness of deep metric learning models has to be improved. In order to avoid model collapse due to excessively hard examples, the existing defenses dismiss the min-max adversarial training, but instead learn from a weak adversary inefficiently. Conversely, we propose Hardness Manipulation to efficiently perturb the training triplet till a specified level of hardness for adversarial training, according to a harder benign triplet or a pseudo-hardness function. It is flexible since regular training and min-max adversarial training are its boundary cases. Besides, Gradual Adversary, a family of pseudo-hardness functions is proposed to gradually increase the specified hardness level during training for a better balance between performance and robustness. Additionally, an Intra-Class Structure loss term among benign and adversarial examples further improves model robustness and efficiency. Comprehensive experimental results suggest that the proposed method, although simple in its form, overwhelmingly outperforms the state-of-the-art defenses in terms of robustness, training efficiency, as well as performance on benign examples.

Deep learning models tend to forget their earlier knowledge while incrementally learning new tasks. This behavior emerges because the parameter updates optimized for the new tasks may not align well with the updates suitable for older tasks. The resulting latent representation mismatch causes forgetting. In this work, we propose ELI: Energy-based Latent Aligner for Incremental Learning, which first learns an energy manifold for the latent representations such that previous task latents will have low energy and the current task latents have high energy values. This learned manifold is used to counter the representational shift that happens during incremental learning. The implicit regularization that is offered by our proposed methodology can be used as a plug-and-play module in existing incremental learning methodologies. We validate this through extensive evaluation on CIFAR-100, ImageNet subset, ImageNet 1k and Pascal VOC datasets. We observe consistent improvement when ELI is added to three prominent methodologies in class-incremental learning, across multiple incremental settings. Further, when added to the state-of-the-art incremental object detector, ELI provides over 5% improvement in detection accuracy, corroborating its effectiveness and complementary advantage to existing art. Code is available at: https://github.com/JosephKJ/ELI.

In this paper, we aim to forecast a future trajectory distribution of a moving agent in the real world, given the social scene images and historical trajectories. Yet, it is a challenging task because the ground-truth distribution is unknown and unobservable, while only one of its samples can be applied for supervising model learning, which is prone to bias. Most recent works focus on predicting diverse trajectories in order to cover all modes of the real distribution, but they may despise the precision and thus give too much credit to unrealistic predictions. To address the issue, we learn the distribution with symmetric cross-entropy using occupancy grid maps as an explicit and scene-compliant approximation to the ground-truth distribution, which can effectively penalize unlikely predictions. In specific, we present an inverse reinforcement learning based multi-modal trajectory distribution forecasting framework that learns to plan by an approximate value iteration network in an end-to-end manner. Besides, based on the predicted distribution, we generate a small set of representative trajectories through a differentiable Transformer-based network, whose attention mechanism helps to model the relations of trajectories. In experiments, our method achieves state-of-the-art performance on the Stanford Drone Dataset and Intersection Drone Dataset.

In this work, we focus on semi-supervised learning for video action detection which utilizes both labeled as well as unlabeled data. We propose a simple end-to-end consistency based approach which effectively utilizes the unlabeled data. Video action detection requires both, action class prediction as well as a spatio-temporal localization of actions. Therefore, we investigate two types of constraints, classification consistency, and spatio-temporal consistency. The presence of predominant background and static regions in a video makes it challenging to utilize spatio-temporal consistency for action detection. To address this, we propose two novel regularization constraints for spatio-temporal consistency; 1) temporal coherency, and 2) gradient smoothness. Both these aspects exploit the temporal continuity of action in videos and are found to be effective for utilizing unlabeled videos for action detection. We demonstrate the effectiveness of the proposed approach on two different action detection benchmark datasets, UCF101-24 and JHMDB-21. In addition, we also show the effectiveness of the proposed approach for video object segmentation on the Youtube-VOS which demonstrates its generalization capability The proposed approach achieves competitive performance by using merely 20% of annotations on UCF101-24 when compared with recent fully supervised methods. On UCF101-24, it improves the score by +8.9% and +11% at 0.5 f-mAP and v-mAP respectively, compared to supervised approach.

The referring video object segmentation task (RVOS) involves segmentation of a text-referred object instance in the frames of a given video. Due to the complex nature of this multimodal task, which combines text reasoning, video understanding, instance segmentation and tracking, existing approaches typically rely on sophisticated pipelines in order to tackle it. In this paper, we propose a simple Transformer-based approach to RVOS. Our framework, termed Multimodal Tracking Transformer (MTTR), models the RVOS task as a sequence prediction problem. Following recent advancements in computer vision and natural language processing, MTTR is based on the realization that video and text can be processed together effectively and elegantly by a single multimodal Transformer model. MTTR is end-to-end trainable, free of text-related inductive bias components and requires no additional mask-refinement post-processing steps. As such, it simplifies the RVOS pipeline considerably compared to existing methods. Evaluation on standard benchmarks reveals that MTTR significantly outperforms previous art across multiple metrics. In particular, MTTR shows impressive +5.7 and +5.0 mAP gains on the A2D-Sentences and JHMDB-Sentences datasets respectively, while processing 76 frames per second. In addition, we report strong results on the public validation set of Refer-YouTube-VOS, a more challenging RVOS dataset that has yet to receive the attention of researchers. The code to reproduce our experiments is available at https://github.com/mttr2021/MTTR

Existing face forgery detectors mainly focus on specific forgery patterns like noise characteristics, local textures, or frequency statistics for forgery detection. This causes specialization of learned representations to known forgery patterns presented in the training set, and makes it difficult to detect forgeries with unknown patterns. In this paper, from a new perspective, we propose a forgery detection framework emphasizing the common compact representations of genuine faces based on reconstruction-classification learning. Reconstruction learning over real images enhances the learned representations to be aware of forgery patterns that are even unknown, while classification learning takes the charge of mining the essential discrepancy between real and fake images, facilitating the understanding of forgeries. To achieve better representations, instead of only using the encoder in reconstruction learning, we build bipartite graphs over the encoder and decoder features in a multi-scale fashion. We further exploit the reconstruction difference as guidance of forgery traces on the graph output as the final representation, which is fed into the classifier for forgery detection. The reconstruction and classification learning is optimized end-to-end. Extensive experiments on large-scale benchmark datasets demonstrate the superiority of the proposed method over state of the arts.

Current methods of multi-person pose estimation typically treat the localization and association of body joints separately. In this paper, we propose the first fully end-to-end multi-person Pose Estimation framework with TRansformers, termed PETR. Our method views pose estimation as a hierarchical set prediction problem and effectively removes the need for many hand-crafted modules like RoI cropping, NMS and grouping post-processing. In PETR, multiple pose queries are learned to directly reason a set of full-body poses. Then a joint decoder is utilized to further refine the poses by exploring the kinematic relations between body joints. With the attention mechanism, the proposed method is able to adaptively attend to the features most relevant to target keypoints, which largely overcomes the feature misalignment difficulty in pose estimation and improves the performance considerably. Extensive experiments on the MS COCO and CrowdPose benchmarks show that PETR plays favorably against state-of-the-art approaches in terms of both accuracy and efficiency. The code and models are available at https://github.com/hikvision-research/opera.

In this paper, we propose an effective and efficient method for Human-Gaze-Target (HGT) detection, i.e., gaze following. Current approaches decouple the HGT detection task into separate branches of salient object detection and human gaze prediction, employing a two-stage framework where human head locations must first be detected and then be fed into the next gaze target prediction sub-network. In contrast, we redefine the HGT detection task as detecting human head locations and their gaze targets, simultaneously. By this way, our method, named Human-Gaze-Target detection TRansformer or HGTTR, streamlines the HGT detection pipeline by eliminating all other additional components. HGTTR reasons about the relations of salient objects and human gaze from the global image context. Moreover, unlike existing two-stage methods that require human head locations as input and can predict only one human's gaze target at a time, HGTTR can directly predict the locations of all people and their gaze targets at one time in an end-to-end manner. The effectiveness and robustness of our proposed method are verified with extensive experiments on the two standard benchmark datasets, GazeFollowing and VideoAttentionTarget. Without bells and whistles, HGTTR outperforms existing state-of-the-art methods by large margins (6.4 mAP gain on GazeFollowing and 10.3 mAP gain on VideoAttentionTarget) with a much simpler architecture.

Recent video and language pretraining frameworks lack the ability to generate sentences. We present Multimodal Video Generative Pretraining (MV-GPT), a new pretraining framework for learning from unlabelled videos which can be effectively used for generative tasks such as multimodal video captioning. Unlike recent video-language pretraining frameworks, our framework trains both a multimodal video encoder and a sentence decoder jointly. To overcome the lack of captions in unlabelled videos, we leverage the future utterance as an additional text source and propose a bidirectional generation objective -- we generate future utterances given the present mulitmodal context, and also the present utterance given future observations. With this objective, we train an encoder-decoder model end-to-end to generate a caption from raw pixels and transcribed speech directly. Our model achieves state-of-the-art performance for multimodal video captioning on four standard benchmarks, as well as for other video understanding tasks such as generative and discriminative VideoQA, video retrieval and action classification.

Generic event boundary detection aims to localize the generic, taxonomy-free event boundaries that segment videos into chunks. Existing methods typically require video frames to be decoded before feeding into the network, which demands considerable computational power and storage space. To that end, we propose a new end-to-end compressed video representation learning for event boundary detection that leverages the rich information in the compressed domain, i.e., RGB, motion vectors, residuals, and the internal group of pictures (GOP) structure, without fully decoding the video. Specifically, we first use the ConvNets to extract features of the I-frames in the GOPs. After that, a light-weight spatial-channel compressed encoder is designed to compute the feature representations of the P-frames based on the motion vectors, residuals and representations of their dependent I-frames. A temporal contrastive module is proposed to determine the event boundaries of video sequences. To remedy the ambiguities of annotations and speed up the training process, we use the Gaussian kernel to preprocess the ground-truth event boundaries. Extensive experiments conducted on the Kinetics-GEBD dataset demonstrate that the proposed method achieves comparable results to the state-of-the-art methods with 4.5x faster running speed.

Although provably robust to translational perturbations, convolutional neural networks (CNNs) are known to suffer from extreme performance degradation when presented at test time with more general geometric transformations of inputs. Recently, this limitation has motivated a shift in focus from CNNs to Capsule Networks (CapsNets). However, CapsNets suffer from admitting relatively few theoretical guarantees of invariance. We introduce a rigourous mathematical framework to permit invariance to any Lie group of warps, exclusively using convolutions (over Lie groups), without the need for capsules. Previous work on group convolutions has been hampered by strong assumptions about the group, which precludes the application of such techniques to common warps in computer vision such as affine and homographic. Our framework enables the implementation of group convolutions over any finite-dimensional Lie group. We empirically validate our approach on the benchmark affine-invariant classification task, where we achieve 30% improvement in accuracy against conventional CNNs while outperforming most CapsNets. As further illustration of the generality of our framework, we train a homography-convolutional model which achieves superior robustness on a homography-perturbed dataset, where CapsNet results degrade.

Generalized zero-shot learning (GZSL) requires a classifier trained on seen classes that can recognize objects from both seen and unseen classes. Due to the absence of unseen training samples, the classifier tends to bias towards seen classes. To mitigate this problem, feature generation based models are proposed to synthesize visual features for unseen classes. However, these features are generated in the visual feature space which lacks of discriminative ability. Therefore, some methods turn to find a better embedding space for the classifier training. They emphasize the inter-class relationships of seen classes, leading the embedding space overfitted to seen classes and unfriendly to unseen classes. Instead, in this paper, we propose an Intra-Class Compactness Enhancement method (ICCE) for GZSL. Our ICCE promotes intra-class compactness with inter-class separability on both seen and unseen classes in the embedding space and visual feature space. By promoting the intra-class relationships but the inter-class structures, we can distinguish different classes with better generalization. Specifically, we propose a Self-Distillation Embedding (SDE) module and a Semantic-Visual Contrastive Generation (SVCG) module. The former promotes intra-class compactness in the embedding space, while the latter accomplishes it in the visual feature space. The experiments demonstrate that our ICCE outperforms the state-of-the-art methods on four datasets and achieves competitive results on the remaining dataset.

In LiDAR-based 3D object detection for autonomous driving, the ratio of the object size to input scene size is significantly smaller compared to 2D detection cases. Overlooking this difference, many 3D detectors directly follow the common practice of 2D detectors, which downsample the feature maps even after quantizing the point clouds. In this paper, we start by rethinking how such multi-stride stereotype affects the LiDAR-based 3D object detectors. Our experiments point out that the downsampling operations bring few advantages, and lead to inevitable information loss. To remedy this issue, we propose Single-stride Sparse Transformer (SST) to maintain the original resolution from the beginning to the end of the network. Armed with transformers, our method addresses the problem of insufficient receptive field in single-stride architectures. It also cooperates well with the sparsity of point clouds and naturally avoids expensive computation. Eventually, our SST achieves state-of-the-art results on the large-scale Waymo Open Dataset. It is worth mentioning that our method can achieve exciting performance (83.8 LEVEL_1 AP on validation split) on small object (pedestrian) detection due to the characteristic of single stride. Our codes will be public soon.

Human pose estimation from single images is a challenging problem that is typically solved by supervised learning. Unfortunately, labeled training data does not yet exist for many human activities since 3D annotation requires dedicated motion capture systems. Therefore, we propose an unsupervised approach that learns to predict a 3D human pose from a single image while only being trained with 2D pose data, which can be crowd-sourced and is already widely available. To this end, we estimate the 3D pose that is most likely over random projections, with the likelihood estimated using normalizing flows on 2D poses. While previous work requires strong priors on camera rotations in the training data set, we learn the distribution of camera angles which significantly improves the performance. Another part of our contribution is to stabilize training with normalizing flows on high-dimensional 3D pose data by first projecting the 2D poses to a linear subspace. We outperform state-of-the-art in unsupervised human pose estimation on the benchmark dataset Human3.6M in all metrics.

A novel algorithm to detect road lanes in the eigenlane space is proposed in this paper. First, we introduce the notion of eigenlanes, which are data-driven descriptors for structurally diverse lanes, including curved, as well as straight, lanes. To obtain eigenlanes, we perform the best rank-M approximation of a lane matrix containing all lanes in a training set. Second, we generate a set of lane candidates by clustering the training lanes in the eigenlane space. Third, using the lane candidates, we determine an optimal set of lanes by developing an anchor-based detection network, called SIIC-Net. Experimental results demonstrate that the proposed algorithm provides excellent detection performance for structurally diverse lanes. Our codes are available at https://github.com/dongkwonjin/Eigenlanes.

Novel contour descriptors, called eigencontours, based on low-rank approximation are proposed in this paper. First, we construct a contour matrix containing all object boundaries in a training set. Second, we decompose the contour matrix into eigencontours via the best rank-M approximation. Third, we represent an object boundary by a linear combination of the M eigencontours. We also incorporate the eigencontours into an instance segmentation framework. Experimental results demonstrate that the proposed eigencontours can represent object boundaries more effectively and more efficiently than existing descriptors in a low-dimensional space. Furthermore, the proposed algorithm yields meaningful performances on instance segmentation datasets.

In this paper, we study a problem of egocentric scene understanding, i.e., predicting depths and surface normals from an egocentric image. Egocentric scene understanding poses unprecedented challenges: (1) due to large head movements, the images are taken from non-canonical viewpoints (i.e., tilted images) where existing models of geometry prediction do not apply; (2) dynamic foreground objects including hands constitute a large proportion of visual scenes. These challenges limit the performance of the existing models learned from large indoor datasets, such as ScanNet and NYUv2, which comprise predominantly upright images of static scenes. We present a multimodal spatial rectifier that stabilizes the egocentric images to a set of reference directions, which allows learning a coherent visual representation. Unlike unimodal spatial rectifier that often produces excessive perspective warp for egocentric images, the multimodal spatial rectifier learns from multiple directions that can minimize the impact of the perspective warp. To learn visual representations of the dynamic foreground objects, we present a new dataset called EDINA (Egocentric Depth on everyday INdoor Activities) that comprises more than 500K synchronized RGBD frames and gravity directions. Equipped with the multimodal spatial rectifier and the EDINA dataset, our proposed method on single-view depth and surface normal estimation significantly outperforms the baselines not only on our EDINA dataset, but also on other popular egocentric datasets, such as First Person Hand Action (FPHA) and EPIC-KITCHENS.