ECCV 2022 Papers

"Data cleaning, architecture, and loss function design are important factors contributing to high-performance face recognition. Previously, the research community tries to improve the performance of each single aspect but failed to present a unified solution on the joint search of the optimal designs for all three aspects.In this paper, we for the first time identify that these aspects are tightly coupled to each other. Optimizing the design of each aspect actually greatly limits the performance and biases the algorithmic design.Specifically, we find that the optimal model architecture or loss function is closely coupled with the data cleaning. To eliminate the bias of single-aspect research and provide an overall understanding of the face recognition model design, we first carefully design the search space for each aspect, then a comprehensive search method is introduced to jointly search optimal data cleaning, architecture, and loss function design.In our framework, we make the proposed comprehensive search as flexible as possible, by using an innovative reinforcement learning based approach.Extensive experiments on million-level face recognition benchmarks demonstrate the effectiveness of our newly-designed search space for each aspect and the comprehensive search. We outperform expert algorithms developed for each single research track by large margins. More importantly, we analyze the difference between our searched optimal design and the independent design of the single factors. We point out that strong models tend to optimize with more difficult training datasets and loss functions. Our empirical study can provide guidance in future research towards more robust face recognition systems."

"With the shift towards on-device deep learning, ensuring a consistent behavior of an AI service across diverse compute platforms becomes tremendously important. Our work tackles the emergent problem of reducing predictive in-consistencies arising as negative flips: test samples that are correctly predicted by a less accurate on-device model, but incorrectly by a more accurate on-cloud one. We introduce REGression constrained Neural Architecture Search (REG-NAS) to design a family of highly accurate models that engender fewer negative flips. REG-NAS consists of two components: (1) A novel architecture constraint that enables a larger on-cloud model to contain all the weights of the smaller on-device one thus maximizing weight sharing. This idea stems from our observation that larger weight sharing among networks leads to similar sample-wise predictions and results in fewer negative flips; (2) A novel search reward that incorporates both Top-1 accuracy and negative flips in the architecture optimization metric. We demonstrate that REG-NAS can successfully find architecture with few negative flips, in three popular architecture search spaces. Compared to the existing state-of-the-art approach [29], REG-NAS leads to 33-48% relative reduction of negative flips."

"Deep learning is pushing the state-of-the-art in many computer vision applications. However, it relies on large annotated data repositories, and capturing the unconstrained nature of the real-world data is yet to be solved. Semi-supervised learning (SSL) complements the annotated training data with a large corpus of unlabeled data to reduce annotation cost. The standard SSL approach assumes unlabeled data are from the same distribution as annotated data. Recently, a more realistic SSL problem, called open-world SSL, is introduced, where the unannotated data might contain samples from unknown classes. In this paper, we propose a novel pseudo-label based approach to tackle SSL in open-world setting. At the core of our method, we utilize sample uncertainty and incorporate prior knowledge about class distribution to generate reliable class-distribution-aware pseudo-labels for unlabeled data belonging to both known and unknown classes. Our extensive experimentation showcases the effectiveness of our approach on several benchmark datasets, where it substantially outperforms the existing state-of-the-art on seven diverse datasets including CIFAR-100 ( 17%), ImageNet-100 ( 5%), and Tiny ImageNet ( 9%). We also highlight the flexibility of our approach in solving novel class discovery task, demonstrate its stability in dealing with imbalanced data, and complement our approach with a technique to estimate the number of novel classes."

"Existing deep learning based HDRTV reconstruction methods assume one kind of tone mapping operators (TMOs) as the degradation procedure to synthesize SDRTV-HDRTV pairs for supervised training. In this paper, we argue that, although traditional TMOs exploit efficient dynamic range compression priors, they have several drawbacks on modeling the realistic degradation: information over-preservation, color bias and possible artifacts, making the trained reconstruction networks hard to generalize well to real-world cases. To solve this problem, we propose a learning-based data synthesis approach to learn the properties of real-world SDRTVs by integrating several tone mapping priors into both network structures and loss functions. In specific, we design a conditioned two-stream network with prior tone mapping results as a guidance to synthesize SDRTVs by both global and local transformations. To train the data synthesis network, we form a novel self-supervised content loss to constraint different aspects of the synthesized SDRTVs at regions with different brightness distributions and an adversarial loss to emphasize the details to be more realistic. To validate the effectiveness of our approach, we synthesize SDRTV-HDRTV pairs with our method and use them to train several HDRTV reconstruction networks. Then we collect two inference datasets containing both labeled and unlabeled real-world SDRTVs, respectively. Experimental results demonstrate that, the networks trained with our synthesized data generalize significantly better to these two real-world datasets than existing solutions."

"Virtual facial avatars will play an increasingly important role in immersive communication, games and the metaverse, and it is therefore critical that they be inclusive. This requires accurate recovery of the albedo, regardless of age, sex, or ethnicity. While significant progress has been made on estimating 3D facial geometry, appearance estimation has received less attention. The task is fundamentally ambiguous because the observed color is a function of albedo and lighting, both of which are unknown. We find that current methods are biased towards light skin tones due to (1) strongly biased priors that prefer lighter pigmentation and (2) algorithmic solutions that disregard the light/albedo ambiguity. To address this, we propose a new evaluation dataset (FAIR) and an algorithm (TRUST) to improve albedo estimation and, hence, fairness. Specifically, we create the first facial albedo evaluation benchmark where subjects are balanced in terms of skin color, and measure accuracy using the Individual Typology Angle (ITA) metric. We then address the light/albedo ambiguity by building on a key observation: the image of the full scene -as opposed to a cropped image of the face- contains important information about lighting that can be used for disambiguation. TRUST regresses facial albedo by conditioning on both the face region and a global illumination signal obtained from the scene image. Our experimental results show significant improvement compared to state-of-the-art methods on albedo estimation, both in terms of accuracy and fairness. The evaluation benchmark and code are available for research purposes at https://trust.is.tue.mpg.de."

"Scene graph generation (SGG) is a fundamental task aimed at detecting visual relations between objects in an image. The prevailing SGG methods require all object classes to be given in the training set. Such a closed setting limits the practical application of SGG. In this paper, we introduce open-vocabulary scene graph generation, a novel, realistic and challenging setting, in which a model is trained on a small set of base object classes but is required to infer relations for unseen target object classes. To this end, we propose a two-step method which firstly pre-trains on large amounts of coarse-grained region-caption data and then leverage two prompt-based techniques to finetune the pre-trained model without updating its parameters. Moreover, our method is able to support inference over completely unseen object classes, which existing methods are incapable of handling. On extensive experiments on three benchmark datasets, Visual Genome, GQA and Open-Image, our method significantly outperforms recent, strong SGG methods on the setting of Ov-SGG, as well as on the conventional closed SGG."

"Open Set Video Anomaly Detection (OpenVAD) aims to identify abnormal events from video data where both known anomalies and novel ones exist in testing. Unsupervised models learned solely from normal videos are applicable to any testing anomalies but suffer from a high false positive rate. In contrast, weakly supervised methods are effective in detecting known anomalies but could fail in an open world. We develop a novel weakly supervised method for the OpenVAD problem by integrating evidential deep learning (EDL) and normalizing flows (NFs) into a multiple instance learning (MIL) framework. Specifically, we propose to use graph neural networks and triplet loss to learn discriminative features for training the EDL classifier, where the EDL is capable of identifying the unknown anomalies by quantifying the uncertainty. Moreover, we develop an uncertainty-aware selection strategy to obtain clean anomaly instances and a NFs module to generate the pseudo anomalies. Our method is superior to existing approaches by inheriting the advantages of both the unsupervised NFs and the weakly-supervised MIL framework. Experimental results on multiple real-world video datasets show the effectiveness of our method."

"Face reconstruction and tracking is a building block of numerous applications in AR/VR, human-machine interaction, as well as medical applications. Most of these applications rely on a metrically correct prediction of the shape, especially, when the reconstructed subject is put into a metrical context (i.e., when there is a reference object of known size). A metrical reconstruction is also needed for any application that measures distances and dimensions of the subject (e.g., to virtually fit a glasses frame). State-of-the-art methods for face reconstruction from a single image are trained on large 2D image datasets in a self-supervised fashion. However, due to the nature of a perspective projection they are not able to reconstruct the actual face dimensions, and even predicting the average human face outperforms some of these methods in a metrical sense. To learn the actual shape of a face, we argue for a supervised training scheme. Since there exists no large-scale 3D dataset for this task, we annotated and unified small- and medium-scale databases. The resulting unified dataset is still a medium-scale dataset with more than 2k identities and training purely on it would lead to overfitting. To this end, we take advantage of a face recognition network pretrained on a large-scale 2D image dataset, which provides distinct features for different faces and is robust to expression, illumination, and camera changes. Using these features, we train our face shape estimator in a supervised fashion, inheriting the robustness and generalization of the face recognition network. Our method, which we call \modellong, outperforms the state-of-the-art reconstruction methods by a large margin, both on current non-metric benchmarks as well as on our metric benchmarks (15% and 24% lower average error on NoW, respectively)."

"We present a method that learns neural shadow fields, which are neural scene representations that are only learnt from the shadows present in the scene. While traditional shape-from-shadow (SfS) algorithms reconstruct geometry from shadows, they assume a fixed scanning setup and fail to generalize to complex scenes. Neural rendering algorithms, on the other hand, rely on photometric consistency between RGB images, but largely ignore physical cues such as shadows, which have been shown to provide valuable information about the scene. We observe that shadows are a powerful cue that can constrain neural scene representations to learn SfS, and even outperform NeRF to reconstruct otherwise hidden geometry. We propose a graphics-inspired differentiable approach to render accurate shadows with volumetric rendering, predicting a shadow map that can be compared to the ground truth shadow. Even with just binary shadow maps, we show that neural rendering can localize the object and estimate coarse geometry. Our approach reveals that sparse cues in images can be used to estimate geometry using differentiable volumetric rendering. Moreover, our framework is highly generalizable and can work alongside existing 3D reconstruction techniques that otherwise only use photometric consistency."

"In this paper, we study a practical space-time video super-resolution (STVSR) problem which aims at generating a high-framerate high-resolution sharp video from a low-framerate low-resolution blurry video. Such problem often occurs when recording a fast dynamic event with a low-framerate and low-resolution camera, and the captured video would suffer from three typical issues: i) motion blur occurs due to object/camera motions during exposure time; ii) motion aliasing is unavoidable when the event temporal frequency exceeds the Nyquist limit of temporal sampling; iii) high-frequency details are lost because of the low spatial sampling rate. These issues can be alleviated by a cascade of three separate sub-tasks, including video deblurring, frame interpolation, and super-resolution, which, however, would fail to capture the spatial and temporal correlations among video sequences. To address this, we propose an interpretable STVSR framework by leveraging both model-based and learning-based methods. Specifically, we formulate STVSR as a joint video deblurring, frame interpolation, and super-resolution problem, and solve it as two sub-problems in an alternate way. For the first sub-problem, we derive an interpretable analytical solution and use it as a Fourier data transform layer. Then, we propose a recurrent video enhancement layer for the second sub-problem to further recover high-frequency details. Extensive experiments demonstrate the superiority of our method in terms of quantitative metrics and visual quality."

"We present a new framework to reconstruct holistic 3D indoor scenes including both room background and indoor objects from single-view images. Existing methods can only produce 3D shapes of indoor objects with limited geometry quality because of the heavy occlusion of indoor scenes. To solve this, we propose an instance-aligned implicit function (InstPIFu) for detailed object reconstruction. Combining with instance-aligned attention module, our method is empowered to decouple mixed local features toward the occluded instances. Additionally, unlike previous methods that simply represents the room background as a 3D bounding box, depth map or a set of planes, we recover the fine geometry of the background via implicit representation. Extensive experiments on the SUN RGB-D, Pix3D, 3D-FUTURE, and 3D-FRONT datasets demonstrate that our method outperforms existing approaches in both background and foreground object reconstruction. Our code and model will be made publicly available."

"Human-Object Interaction (HOI) detection is a core task for high-level image understanding. Recently, Detection Transformer (DETR)-based HOI detectors have become popular due to their superior performance and efficient structure. However, these approaches typically adopt fixed HOI queries for all testing images, which is vulnerable to the location change of objects in one specific image. Accordingly, in this paper, we propose to enhance DETR’s robustness by mining hard-positive queries, which are forced to make correct predictions using partial visual cues. First, we explicitly compose hard-positive queries according to the ground-truth (GT) position of labeled human-object pairs for each training image. Specifically, we shift the GT bounding boxes of each labeled human-object pair so that the shifted boxes cover only a certain portion of the GT ones. We encode the coordinates of the shifted boxes for each labeled human-object pair into an HOI query. Second, we implicitly construct another set of hard-positive queries by masking the top scores in cross-attention maps of the decoder layers. The masked attention maps then only cover partial important cues for HOI predictions. Finally, an alternate strategy is proposed that efficiently combines both types of hard queries. In each iteration, both DETR’s learnable queries and one selected type of hard-positive queries are adopted for loss computation. Experimental results show that our proposed approach can be widely applied to existing DETR-based HOI detectors. Moreover, we consistently achieve state-of-the-art performance on three benchmarks: HICO-DET, V-COCO, and HOI-A."

"We present a unified method, termed Unicorn, that can simultaneously solve four tracking problems (SOT, MOT, VOS, MOTS) with a single network using the same model parameters. Due to the fragmented definitions of the object tracking problem itself, most existing trackers are developed to address a single or part of tasks and overspecialize on the characteristics of specific tasks. By contrast, Unicorn provides a unified solution, adopting the same input, backbone, embedding, and head across all tracking tasks. For the first time, we accomplish the great unification of the tracking network architecture and learning paradigm. Unicorn performs on-par or better than its task-specific counterparts in 8 tracking datasets, including LaSOT, TrackingNet, MOT17, BDD100K, DAVIS16-17, MOTS20, and BDD100K MOTS. We believe that Unicorn will serve as a solid step towards the general vision model. Code is available at https://github.com/MasterBin-IIAU/Unicorn."

"Tracking in 3D scenes is gaining momentum because of its numerous applications in robotics, autonomous driving, and scene understanding. Currently, 3D tracking is limited to specific model-based approaches involving point clouds, which impedes 3D trackers from applying in natural 3D scenes. RGBD sensors provide a more reasonable and acceptable solution for 3D object tracking due to their readily available synchronised color and depth information. Thus, in this paper, we investigate a novel problem: is it possible to track a generic (class-agnostic) 3D object in RGBD videos and predict 3D bounding boxes of the object of interest? To inspire further research on this topic, we newly construct a standard benchmark for generic 3D object tracking, ‘Track-it-in-3D’, which contains 300 RGBD video sequences with dense 3D annotations and corresponding evaluation protocols. Furthermore, we propose an effective tracking baseline to estimate 3D bounding boxes for arbitrary objects in RGBD videos, by fusing appearance and spatial information effectively. The dataset and codes will be publicly available."

"With the rapid development of mobile devices, modern widely-used mobile phones typically allow users to capture 4K resolution (i.e., ultra-high-definition) images. However, for image demoiréing, a challenging task in low-level vision, existing works are generally carried out on low-resolution or synthetic images. Hence, the effectiveness of these methods on 4K resolution images is still unknown. In this paper, we explore moiré pattern removal for ultra-high-definition images. To this end, we propose the first ultra-high-definition demoiréing dataset (UHDM), which contains 5,000 real-world 4K resolution image pairs, and conduct a benchmark study on current state-of-the-art methods. Further, we present an efficient baseline model ESDNet for tackling 4K moiré images, wherein we build a semantic-aligned scale-aware module to address the scale variation of moiré patterns. Extensive experiments manifest the effectiveness of our approach, which outperforms state-of-the-art methods by a large margin while being much more lightweight. Code and dataset are available at https://xinyu-andy.github.io/uhdm-page."

"Self-supervision has emerged as a propitious method for visual representation learning after the recent paradigm shift from handcrafted pretext tasks to instance-similarity based approaches. Most state-of-the-art methods enforce similarity between various augmentations of a given image, while some methods additionally use contrastive approaches to explicitly ensure diverse representations. While these approaches have indeed shown promising direction, they require a significantly larger number of training iterations when compared to the supervised counterparts. In this work, we explore reasons for the slow convergence of these methods, and further propose to strengthen them using well-posed auxiliary tasks that converge significantly faster, and are also useful for representation learning. The proposed method utilizes the task of rotation prediction to improve the efficiency of existing state-of-the-art methods. We demonstrate significant gains in performance using the proposed method on multiple datasets, specifically for lower training epochs."

"Vision Transformer (ViT), as a powerful alternative to Convolutional Neural Network (CNN), has received much attention. Recent work showed that ViTs are also vulnerable to adversarial examples like CNNs. To build robust ViTs, an intuitive way is to apply adversarial training since it has been shown as one of the most effective ways to accomplish robust CNNs. However, one major limitation of adversarial training is its heavy computational cost. The self-attention mechanism adopted by ViTs is a computationally intense operation whose expense increases quadratically with the number of input patches, making adversarial training on ViTs even more time-consuming. In this work, we first comprehensively study fast adversarial training on a variety of vision transformers and illustrate the relationship between the efficiency and robustness. Then, to expediate adversarial training on ViTs, we propose an efficient Attention Guided Adversarial Training mechanism. Specifically, relying on the specialty of self-attention, we actively remove certain patch embeddings of each layer with an attention-guided dropping strategy during adversarial training. The slimmed self-attention modules accelerate the adversarial training on ViTs significantly. With only 65\% of the fast adversarial training time, we match the state-of-the-art results on the challenging ImageNet benchmark."

"Trojan attacks on deep neural networks are both dangerous and surreptitious. Over the past few years, Trojan attacks have advanced from using only a single input-agnostic trigger and targeting only one class to using multiple, input-specific triggers and targeting multiple classes. However, Trojan defenses have not caught up with this development. Most defense methods still make out-of-date assumptions about Trojan triggers and target classes, thus, can be easily circumvented by modern Trojan attacks. To deal with this problem, we propose two novel ""filtering"" defenses called Variational Input Filtering (VIF) and Adversarial Input Filtering (AIF) which leverage lossy data compression and adversarial learning respectively to effectively purify all potential Trojan triggers in the input at run time without making assumptions about the number of triggers/target classes or the input dependence property of triggers. In addition, we introduce a new defense mechanism called ""Filtering-then-Contrasting"" (FtC) which helps avoid the drop in classification accuracy on clean data caused by ""filtering"", and combine it with VIF/AIF to derive new defenses of this kind. Extensive experimental results and ablation studies show that our proposed defenses significantly outperform well-known baseline defenses in mitigating five advanced Trojan attacks including two recent state-of-the-art while being quite robust to small amounts of training data and large-norm triggers."

"Detection transformers have achieved competitive performance on the sample-rich COCO dataset. However, we show most of them suffer from significant performance drops on small-size datasets, like Cityscapes. In other words, the detection transformers are generally data-hungry. To tackle this problem, we empirically analyze the factors that affect data efficiency, through a step-by-step transition from a data-efficient RCNN variant to the representative DETR. The empirical results suggest that sparse feature sampling from local image areas holds the key. Based on this observation, we alleviate the data-hungry issue of existing detection transformers by simply alternating how key and value sequences are constructed in the cross-attention layer, with minimum modifications to the original models. Besides, we introduce a simple yet effective label augmentation method to provide richer supervision and improve data efficiency. Experiments show that our method can be readily applied to different detection transformers and improve their performance on both small-size and sample-rich datasets."