WACV 2023 Papers

A promising direction for recovering the lost information in low-resolution headshot images is utilizing a set of high-resolution exemplars from the same identity. Complementary images in the reference set can improve the generated headshot quality across many different views and poses. However, it is challenging to make the best use of multiple exemplars: the quality and alignment of each exemplar cannot be guaranteed. Using low-quality and mismatched images as references will impair the output results. To overcome these issues, we propose the efficient Headshot Image Super-Resolution with Multiple Exemplars network (HIME) method. Compared with previous methods, our network can effectively handle the misalignment between the input and the reference without requiring facial priors and learn the aggregated reference set representation in an end-to-end manner. Furthermore, to reconstruct more detailed facial features, we propose a correlation loss that provides a rich representation of the local texture in a controllable spatial range. Experimental results demonstrate that the proposed framework not only has significantly fewer computation cost than recent exemplar-guided methods but also achieves better qualitative and quantitative performance.

The current success of modern visual reasoning systems is arguably attributed to cross-modality attention mechanisms. However, in deliberative reasoning such as in VQA, attention is unconstrained at each step, and thus may serve as a statistical pooling mechanism rather than a semantic operation intended to select information relevant to inference. This is because at training time, attention is only guided by a very sparse signal (i.e. the answer label) at the end of the inference chain. This causes the cross-modality attention weights to deviate from the desired visual-language bindings. To rectify this deviation, we propose to guide the attention mechanism using explicit linguistic-visual grounding. This grounding is derived by connecting structured linguistic concepts in the query to their referents among the visual objects. Here we learn the grounding from the pairing of questions and images alone, without the need for answer annotation or external grounding supervision. This grounding guides the attention mechanism inside VQA models through a duality of mechanisms: pre-training attention weight calculation and directly guiding the weights at inference time on a case-by-case basis. The resultant algorithm is capable of probing attention-based reasoning models, injecting relevant associative knowledge, and regulating the core reasoning process. This scalable enhancement improves the performance of VQA models, fortifies their robustness to limited access to supervised data, and increases interpretability.

Existing deep interactive colorization models have focused on ways to utilize various types of interactions, such as point-wise color hints, scribbles, or natural-language texts, as methods to reflect a user's intent at runtime. However, another approach, which actively informs the user of the most effective regions to give hints for sketch image colorization, has been under-explored. This paper proposes a novel model-guided deep interactive colorization framework that reduces the required amount of user interactions, by prioritizing the regions in a colorization model. Our method, called GuidingPainter, prioritizes these regions where the model most needs a color hint, rather than just relying on the user's manual decision on where to give a color hint. In our extensive experiments, we show that our approach outperforms existing interactive colorization methods in terms of the conventional metrics, such as PSNR and FID, and reduces required amount of interactions.

This paper presents a framework for jointly grounding objects that follow certain semantic relationship constraints given in a scene graph. A typical natural scene contains several objects, often exhibiting visual relationships of varied complexities between them. These inter-object relationships provide strong contextual cues towards improving grounding performance compared to a traditional object query-only-based localization task. A scene graph is an efficient and structured way to represent all the objects and their semantic relationships in the image. In an attempt towards bridging these two modalities representing scenes and utilizing contextual information for improving object localization, we rigorously study the problem of grounding scene graphs on natural images. To this end, we propose a novel graph neural network-based approach referred to as Visio-Lingual Message PAssing Graph Neural Network (VL-MPAG Net). In VL-MPAG Net, we first construct a directed graph with object proposals as nodes and an edge between a pair of nodes representing a plausible relation between them. Then a three-step inter-graph and intra-graph message passing is performed to learn the context-dependent representation of the proposals and query objects. These object representations are used to score the proposals to generate object localization. The proposed method significantly outperforms the baselines on four public datasets.

Existing deep learning approaches for person re-identification (Re-ID) mostly rely on large-scale and well-annotated training data. However, human-annotated labels are prone to label noise in real-world applications. Previous person Re-ID works mainly focus on random label noise, which doesn't properly reflect the characteristic of label noise in practical human-annotated process. In this work, we find the visual ambiguity noise is more common and reasonable noise assumption in annotation of person Re-ID. To handle the kind of noise, we propose a simple and effective robust person Re-ID framework, namely Graph-Based Self-Learning (GBSL), to iteratively learn discriminative representation and rectify noisy labels with limited annotated samples for each identity. Meanwhile, considering the practical annotation process in person Re-ID, we further extend the visual ambiguity noise assumption and propose a type of more practical label noise in person Re-ID, namely the tracklet-level label noise (TLN). Without modifying network architecture or loss function, our approach significantly improves the robustness against label noise of the Re-ID system. Our model obtains competitive performance with training data corrupted by various types of label noise and outperforms the existing methods for robust Re-ID on public benchmarks.

The majority of uncertainty quantification methods for deep object detectors are based on the network output, such as sampling strategies like Monte-Carlo dropout or deep ensembles with straight-forward transfers to object detection. Here, we study gradient-based uncertainty features for object detection. We show that they contain information orthogonal to that of common, output-based uncertainty approximation methods. Meta classification and meta regression are used to produce confidence estimates using gradient features and other methods which are applicable to numerous object detection architectures. Our results show that gradient uncertainty itself performs on par with stateof-the-art methods across different detectors and datasets. We find that combined meta classifiers outperform standalone models. This suggests that sampling strategies may be supplemented by gradient-based uncertainty to obtain improved confidences, contributing to the probabilistic reliability of object detectors in down-stream applications.

Videos shot by laymen using hand-held cameras contain undesirable shaky motion. Estimating the global motion between successive frames, in a manner not influenced by moving objects, is central to many video stabilization techniques, but poses significant challenges. A large body of work uses 2D affine transformations or homography for the global motion. However, in this work, we introduce a more general representation scheme, which adapts any existing optical flow network to ignore the moving objects and obtain a spatially smooth approximation of the global motion between video frames. We achieve this by a knowledge distillation approach, where we first introduce a low pass filter module into the optical flow network to constrain the predicted optical flow to be spatially smooth. This becomes our student network, named as GLOBALFLOWNET. Then, using the original optical flow network as the teacher network, we train the student network using a robust loss function. Given a trained GLOBALFLOWNET, we stabilize videos using a two stage process. In the first stage, we correct the instability in affine parameters using a quadratic programming approach constrained by a user-specified cropping limit to control loss of field of view. In the second stage, we stabilize the video further by smoothing global motion parameters, expressed using small number of discrete cosine transform coefficients. In extensive experiments on a variety of different videos, our technique outperforms state of the art techniques in terms of subjective quality and different quantitative measures of video stability. Additionally, we present a new measure for evaluation of video stabilization based on the flow generated by GLOBALFLOWNET and argue that it is based on a more general motion model in contrast to the affine motion model on which most existing measures are based. The source code is publicly available at https://github.com/GlobalFlowNet/GlobalFlowNet

The downstream accuracy of self-supervised methods is tightly linked to the proxy task solved during training and the quality of the gradients extracted from it. Richer and more meaningful gradients updates are key to allow self-supervised methods to learn better and in a more efficient manner. In a typical self-distillation framework, the representation of two augmented images are enforced to be coherent at the global level. Nonetheless, incorporating local cues in the proxy task can be beneficial and improve the model accuracy on downstream tasks. This leads to a dual objective in which, on the one hand, coherence between global-representations is enforced and on the other, coherence between local-representations is enforced. Unfortunately, an exact correspondence mapping between two sets of local-representations does not exist making the task of matching local-representations from one augmentation to another non-trivial. We propose to leverage the spatial information in the input images to obtain geometric matchings and compare this geometric approach against previous methods based on similarity matchings. Our study shows that not only 1) geometric matchings perform better than similarity based matchings in low-data regimes but also 2) that similarity based matchings are highly hurtful in low-data regimes compared to the vanilla baseline without local self-distillation. The code is available at https://github.com/tileb1/global-local-self-distillation.

Many online action prediction models observe complete frames to locate and attend to informative subregions in the frames called glimpses and recognize an ongoing action based on global and local information. However, in applications with constrained resources, an agent may not be able to observe the complete frame, yet must still locate useful glimpses to predict an incomplete action based on local information only. In this paper, we develop Glimpse Transformers (GliTr), which observe only narrow glimpses at all times, thus predicting an ongoing action and the following most informative glimpse location based on the partial spatiotemporal information collected so far. In the absence of a ground truth for the optimal glimpse locations for action recognition, we train GliTr using a novel spatiotemporal consistency objective: We require GliTr to attend to the glimpses with features similar to the corresponding complete frames (i.e. spatial consistency) and the resultant class logits at time t equivalent to the ones predicted using whole frames up to t (i.e. temporal consistency). Inclusion of our proposed consistency objective yields 10% higher accuracy on the Something-Something-v2 (SSv2) dataset than the baseline cross-entropy objective. Overall, despite observing only 33% of the total area per frame, GliTr achieves 53.02% and 93.91% accuracy on the SSv2 and Jester datasets, respectively.

Reference-guided image inpainting restores image pixels by leveraging the content from another single reference image. The primary challenge is how to precisely place the pixels from the reference image into the hole region. Therefore, understanding the 3D geometry that relates pixels between two views is a crucial step towards building a better model. Given the complexity of handling various types of reference images, we focus on the scenario where the images are captured by freely moving the same camera around. Compared to the previous work, we propose a principled approach that does not make heuristic assumptions about the planarity of the scene. We leverage a monocular depth estimate and predict relative pose between cameras, then align the reference image to the target by a differentiable 3D reprojection and a joint optimization of relative pose and depth map scale and offset. Our approach achieves state-of-the-art performance on both RealEstate10K and MannequinChallenge dataset with large baselines, complex geometry and extreme camera motions. We experimentally verify our approach is also better at handling large holes.

3D LiDAR sensors are indispensable for the robust vision of autonomous mobile robots. However, deploying LiDAR-based perception algorithms often fails due to a domain gap from the training environment, such as inconsistent angular resolution and missing properties. Existing studies have tackled the issue by learning inter-domain mapping, while the transferability is constrained by the training configuration and the training is susceptible to peculiar lossy noises called ray-drop. To address the issue, this paper proposes a generative model of LiDAR range images applicable to the data-level domain transfer. Motivated by the fact that LiDAR measurement is based on point-by-point range imaging, we train an implicit image representation-based generative adversarial networks along with a differentiable ray-drop effect. We demonstrate the fidelity and diversity of our model in comparison with the point-based and image-based state-of-the-art generative models. We also showcase upsampling and restoration applications. Furthermore, we introduce a Sim2Real application for LiDAR semantic segmentation. We demonstrate that our method is effective as a realistic ray-drop simulator and outperforms state-of-the-art methods.

Color is a critical design factor for web pages, affecting important factors such as viewer emotions and the overall trust and satisfaction of a website. Effective coloring requires design knowledge and expertise, but if this process could be automated through data-driven modeling, efficient exploration and alternative workflows would be possible. However, this direction remains underexplored due to the lack of a formalization of the web page colorization problem, datasets, and evaluation protocols. In this work, we propose a new dataset consisting of e-commerce mobile web pages in a tractable format, which are created by simplifying the pages and extracting canonical color styles with a common web browser. The web page colorization problem is then formalized as a task of estimating plausible color styles for a given web page content with a given hierarchical structure of the elements. We present several Transformer-based methods that are adapted to this task by prepending structural message passing to capture hierarchical relationships between elements. Experimental results, including a quantitative evaluation designed for this task, demonstrate the advantages of our methods over statistical and image colorization methods. The code is available at https://github.com/CyberAgentAILab/webcolor.

Existing domain adaptation literature comprises multiple techniques that align the labeled source and unlabeled target domains at different stages, and predict the target labels. In a source-free domain adaptation setting, the source data is not available for alignment. We present a source-free generative paradigm that captures the relations between the source categories and enforces them onto the unlabeled target data, thereby circumventing the need for source data without introducing any new hyper-parameters. The adaptation is performed through the adversarial alignment of the posterior probabilities of the source and target categories. The proposed approach demonstrates competitive performance against other source-free domain adaptation techniques and can also be used for source-present settings.

We present a particle-based neural garment simulator (dubbed as GarSim) that can simulate template garments on the target arbitrary body poses. Existing learning-based methods majorly work for specific garment type (e.g. t-shirt, skirt, etc) or garment topology, and needs retraining for a new type of garment. Similarly, some methods focus on a particular fabric, body shape, and pose. To circumvent these limitations, our method fundamentally learns the physical dynamics of the garment vertices conditioned on underlying body shape, motion, and fabric properties to generalize across garment types, topology, and fabric along with different body shape and pose. In particular, we represent the garment as a graph, where the nodes represent the physical state of the garment vertices, and the edges represent the relation between the two nodes. The nodes and edges of the garment graph encode various properties of garments and the human body to compute the dynamics of the vertices through a learned message-passing. Learning of such dynamics of the garment vertices conditioned on underlying body motion and fabric properties enables our method to be trained simultaneously for multiple types of garments (e.g., tops, skirts, etc) with arbitrary mesh resolutions, varying topologies, and fabric properties. Our experimental results show that GarSim with less amount of training data not only outperforms the SOTA methods on challenging CLOTH3D dataset both qualitatively and quantitatively, but also works reliably well on the unseen poses obtained from YouTube videos, and give satisfactory results on unseen cloth types which were not present during the training.

In person search, we aim to localize a query person from one scene in other gallery scenes. The cost of this search operation is dependent on the number of gallery scenes, making it beneficial to reduce the pool of likely scenes. We describe and demonstrate the Gallery Filter Network (GFN), a novel module which can efficiently discard gallery scenes from the search process, and benefit scoring for persons detected in remaining scenes. We show that the GFN is robust under a range of different conditions by testing on different retrieval sets, including cross-camera, occluded, and low-resolution scenarios. In addition, we develop the base SeqNeXt person search model, which improves and simplifies the original SeqNet model. We show that the SeqNeXt+GFN combination yields significant performance gains over other state-of-the-art methods on the standard PRW and CUHK-SYSU person search datasets. To aid experimentation for this and other models, we provide standardized tooling for the data processing and evaluation pipeline typically used for person search research.

Gait recognition, which identifies individuals based on their walking patterns, is an important biometric technique since it can be observed from a distance and does not require the subject's cooperation. Recognizing a person's gait is difficult because of the appearance variants in human silhouette sequences produced by varying viewing angles, carrying objects, and clothing. Recent research has produced a number of ways for coping with these variants. In this paper, we present the usage of inferring 3-D body shapes distilled from limited images, which are, in principle, invariant to the specified variants. Inference of 3-D shape is a difficult task, especially when only silhouettes are provided in a dataset. We provide a method for learning 3-D body inference from silhouettes by transferring knowledge from 3-D shape prior from RGB photos. We use our method on multiple existing state-of-the-art gait baselines and obtain consistent improvements for gait identification on two public datasets, CASIA-B and OUMVLP, on several variants and settings, including a new setting of novel views not seen during training.

While point cloud semantic segmentation is a significant task in 3D scene understanding, this task demands a time-consuming process of fully annotating labels. To address this problem, recent studies adopt a weakly supervised learning approach under the sparse annotation. Different from the existing studies, this study aims to reduce the epistemic uncertainty measured by the entropy for a precise semantic segmentation. We propose the graphical information gain based attention network called GaIA, which alleviates the entropy of each point based on the reliable information. The graphical information gain discriminates the reliable point by employing relative entropy between target point and its neighborhoods. We further introduce anchor-based additive angular margin loss, ArcPoint. The ArcPoint optimizes the unlabeled points containing high entropy towards semantically similar classes of the labeled points on hypersphere space. Experimental results on S3DIS and ScanNet-v2 datasets demonstrate our framework outperforms the existing weakly supervised methods.

Learning to detect automatic anomalies in production plants remains a machine learning challenge. Since anomalies by definition cannot be learned, their detection must rely on a very accurate "normality model". To this aim, we introduce here a global-to-local Gaussian model for neural network features, learned from a set of normal images. This probabilistic model enables unsupervised anomaly detection. A global Gaussian mixture model of the features is first learned using all available features from normal data. This global Gaussian mixture model is then localized by an adaptation of the K-MLE algorithm, which learns a spatial weight map for each Gaussian. These weights are then used instead of the mixture weights to detect anomalies. This method enables precise modeling of complex data, even with limited data. Applied on WideResnet50-2 features, our approach outperforms the previous state of the art on the MVTec dataset, particularly on the object category. It is robust to perturbations that are frequent in production lines, such as imperfect alignment, and is on par in terms of memory and computation time with the previous state of the art.

Applications based on image retrieval require editing and associating in intermediate spaces that are representative of the high-level concepts like objects and their relationships rather than dense, pixel-level representations like RGB images or semantic-label maps. We focus on one such representation, scene graphs, and propose a novel scene expansion task where we enrich an input seed graph by adding new nodes (objects) and the corresponding relationships. To this end, we formulate scene graph expansion as a sequential prediction task involving multiple iterations of first predicting a new node and then predicting the set of relationships between the newly predicted node and previously chosen nodes in the graph. We propose and evaluate a sequencing strategy that retains the clustering patterns amongst nodes. In addition, we leverage external knowledge to train our graph generation model, enabling greater generalization of node predictions. Due to the inefficiency of existing maximum mean discrepancy (MMD) based metrics standard for graph generation problems, we design novel metrics that comprehensively evaluate different aspects of node and relation predictions. We conduct extensive experiments on Visual Genome and VRD datasets to evaluate the expanded scene graphs using the standard MMD based metrics and our proposed metrics. We observe that the graphs generated by our method, GEMS, better represent the real distribution of the scene graphs compared with baseline methods like GraphRNN.