WACV 2024 Papers

The progress in the generation of synthetic images has made it crucial to assess their quality. While several metrics have been proposed to assess the rendering of images, it is crucial for Text-to-Image (T2I) models, which generate images based on a prompt, to consider additional aspects such as to which extent the generated image matches the important content of the prompt. Moreover, although the generated images usually result from a random starting point, the influence of this one is generally not considered. In this article, we propose a new metric based on prompt templates to study the alignment between the content specified in the prompt and the corresponding generated images. It allows us to better characterize the alignment in terms of the type of the specified objects, their number, and their color. We conducted a study on several recent T2I models about various aspects. An additional interesting result we obtained with our approach is that image quality can vary drastically depending on the noise used as a seed for the images. We also quantify the influence of the number of concepts in the prompt, their order as well as their (color) attributes. Finally, our method allows us to identify some seeds that produce better images than others, opening novel directions of research on this understudied topic.

Cross-modal Steganography is the practice of concealing secret signals in publicly available cover signals (distinct from the modality of the secret signals) unobtrusively. While previous approaches primarily concentrated on concealing a relatively small amount of information, we propose THInImg, which manages to hide lengthy audio data (and subsequently decode talking head video) inside an identity image by leveraging the properties of human face, which can be effectively utilized for covert communication, transmission and copyright protection. THInImg consists of two parts: the encoder and decoder. Inside the encoder-decoder pipeline, we introduce a novel architecture that substantially increase the capacity of hiding audio in images. Moreover, our framework can be extended to iteratively hide multiple audio clips into an identity image, offering multiple levels of control over permissions. We conduct extensive experiments to prove the effectiveness of our method, demonstrating that THInImg can present up to 80 seconds of high quality talking-head video (including audio) in an identity image with 160x160 resolution.

Recent work in Neural Fields (NFs) learn 3D representations from class-specific single view image collections. However, they are unable to reconstruct the input data preserving high-frequency details. Further, these methods do not disentangle appearance from geometry and hence are not suitable for tasks such as texture transfer and editing. In this work, we propose TEGLO (Textured EG3D-GLO) for learning 3D representations from single view in-the-wild image collections for a given class of objects. We accomplish this by training a conditional Neural Radiance Field (NeRF) without any explicit 3D supervision. We equip our method with editing capabilities by creating a dense correspondence mapping to a 2D canonical space. We demonstrate that such mapping enables texture transfer and texture editing without requiring meshes with shared topology. Our key insight is that by mapping the input image pixels onto the texture space we can achieve near perfect reconstruction (>74 dB PSNR at 1024^2 resolution). Our formulation allows for high quality 3D consistent novel view synthesis with high-frequency details even at megapixel image resolutions.

In class-incremental learning (CIL), when deep neural networks learn new classes, their recognition performance in old classes will drop significantly. This phenomenon is widely known as catastrophic forgetting. To alleviate catastrophic forgetting, existing methods store a small portion of old class data with a memory buffer and replay it while learning new classes. These methods suffer from a severe imbalance problem between old and new classes. In this paper, we discover that the imbalance problem in CIL makes it difficult to preserve the feature relation of old classes and hard to learn the feature relation between old and new classes. To mitigate the above two issues, we design a triplet contrastive preserving (TCP) loss to preserve old knowledge, and propose an asymmetric augmented contrastive learning (A2CL) method to learn new classes. Comprehensive experiments demonstrate the effectiveness of our method, which increases the average accuracies by 1.26% and 0.95% on CIFAR-100 and ImageNet. Especially under smaller memory buffer settings where the imbalance problem is more severe, our method can surpass the baselines by a large margin (up to 3.2%). We also show that TCP can be easily plugged into other methods and further improve their performance.

Due to the steadily rising amount of valuable goods in supply chains, tampering detection for parcels is becoming increasingly important. In this work, we focus on the use-case last-mile delivery, where only a single RGB image is taken and compared against a reference from an existing database to detect potential appearance changes that indicate tampering. We propose a tampering detection pipeline that utilizes keypoint detection to identify the eight corner points of a parcel. This permits applying a perspective transformation to create normalized fronto-parallel views for each visible parcel side surface. These viewpoint-invariant parcel side surface representations facilitate the identification of signs of tampering on parcels within the supply chain, since they reduce the problem to parcel side surface matching with pair-wise appearance change detection. Experiments with multiple classical and deep learning-based change detection approaches are performed on our newly collected TAMpering detection dataset for PARcels, called TAMPAR. We evaluate keypoint and change detection separately, as well as in a unified system for tampering detection. Our evaluation shows promising results for keypoint (Keypoint AP 75.76) and tampering detection (81% accuracy, F1-Score 0.83) on real images. Furthermore, a sensitivity analysis for tampering types, lens distortion and viewing angles is presented. Code and dataset are available at https://a-nau.github.io/tampar.

Conditioned diffusion models have demonstrated state-of-the-art text-to-image synthesis capacity. Recently, most works focus on synthesizing independent images; While for real-world applications, it is common and necessary to generate a series of coherent images for story-stelling. In this work, we mainly focus on story visualization and continuation tasks and propose AR-LDM, a latent diffusion model auto-regressively conditioned on history captions and generated images. Moreover, AR-LDM can generalize to new characters through adaptation. To our best knowledge, this is the first work successfully leveraging diffusion models for coherent visual story synthesizing. It also extends the text-conditioned method to multimodal conditioning. Quantitative results show that AR-LDM achieves SoTA FID scores on PororoSV, FlintstonesSV, and the adopted challenging dataset VIST containing natural images. Large-scale human evaluations show that AR-LDM has superior performance in terms of quality, relevance, and consistency.

We address the task of in-the-wild human figure synthesis, where the primary goal is to synthesize a full body given any region in any image. In-the-wild human figure synthesis has long been a challenging and under-explored task, where current methods struggle to handle extreme poses, occluding objects, and complex backgrounds. Our main contribution is TriA-GAN, a keypoint-guided GAN that can synthesize Anyone, Anywhere, in Any given pose. Key to our method is projected GANs combined with a well-crafted training strategy, where our simple generator architecture can successfully handle the challenges of in-the-wild full-body synthesis. We show that TriA-GAN significantly improves over previous in-the-wild full-body synthesis methods, all while requiring less conditional information for synthesis (keypoints v.s. DensePose). Finally, we show that the latent space of TriA-GAN is compatible with standard unconditional editing techniques, enabling text-guided editing of generated human figures.

Synthetic datasets, recognized for their cost effectiveness, play a pivotal role in advancing computer vision tasks and techniques. However, when it comes to remote sensing image processing, the creation of synthetic datasets becomes challenging due to the demand for larger-scale and more diverse 3D models. This complexity is compounded by the difficulties associated with real remote sensing datasets, including limited data acquisition and high annotation costs, which amplifies the need for high-quality synthetic alternatives. To address this, we present SyntheWorld, a synthetic dataset unparalleled in quality, diversity, and scale. It includes 40,000 images with submeter-level pixels and fine-grained land cover annotations of eight categories, and it also provides 40,000 pairs of bitemporal image pairs with building change annotations for building change detection. We conduct experiments on multiple benchmark remote sensing datasets to verify the effectiveness of SyntheWorld and to investigate the conditions under which our synthetic data yield advantages. The dataset is available at https://github.com/JTRNEO/SyntheWorld.

Generative Adversarial Networks (GANs) can generate hyperrealistic face images of synthetic identities based on a latent understanding of real images from a large training set. Despite their proficiency, the term "synthetic identity" remains ambiguous, and the uniqueness of the faces GANs produce is rarely assessed. Recent studies have found that identities from the training data can unintentionally appear in the faces generated by StyleGAN2, but the cause of this phenomenon is unclear. In this work, we propose a novel framework, SynthProv, that utilizes the improved interpolation ability of StyleGAN2 latent space and employs image composition to analyze leakage. This is the first method that goes beyond detection and traces the source or provenance of constituent identity signals in the generated image. Experiments show that SynthProv succeeds in both detection and provenance tasks using multiple matching strategies. We identify identities from FFHQ and CelebA-HQ training datasets with the highest leakage into the latent space as "leaking reals". Analyzing latent space behavior to evaluate generative model privacy via leakage is an important research direction, as undetected leaking reals pose a significant threat to training data privacy. Our code is available at https://github.com/jaisidhsingh/SynthProv

In recent years, continuous latent space (CLS) and discrete latent space (DLS) deep learning models have been proposed for medical image analysis for improved performance. However, these models encounter distinct challenges. CLS models capture intricate details but often lack interpretability in terms of structural representation and robustness due to their emphasis on low-level features. Conversely, DLS models offer interpretability, robustness, and the ability to capture coarse-grained information thanks to their structured latent space. However, DLS models have limited efficacy in capturing fine-grained details. To address the limitations of both DLS and CLS models, we propose SynergyNet, a novel bottleneck architecture designed to enhance existing encoder-decoder segmentation frameworks. SynergyNet seamlessly integrates discrete and continuous representations to harness complementary information and successfully preserves both fine and coarsegrained details in the learned representations. Our extensive experiment on multi-organ segmentation and cardiac datasets demonstrates that SynergyNet outperforms other state of the art methods including TransUNet: dice scores improving by 2.16%, and Hausdorff scores improving by 11.13%, respectively. When evaluating skin lesion and brain tumor segmentation datasets, we observe a remarkable improvements of 1.71% in Intersection-overUnion scores for skin lesion segmentation and of 8.58% for brain tumor segmentation. Our innovative approach paves the way for enhancing the overall performance and capabilities of deep learning models in the critical domain of medical image analysis.

Recently, the fundamental problem of unsupervised domain adaptation (UDA) on 3D point clouds has been motivated by a wide variety of applications in robotics, virtual reality, and scene understanding, to name a few. The point cloud data acquisition procedures manifest themselves as significant domain discrepancies and geometric variations among both similar and dissimilar classes. The standard domain adaptation methods developed for images do not directly translate to point cloud data because of their complex geometric nature. To address this challenge, we leverage the idea of multimodality and alignment between distributions. We propose a new UDA architecture for point cloud classification that benefits from multimodal contrastive learning to get better class separation in both domains individually. Further, the use of optimal transport (OT) aims at learning source and target data distributions jointly to reduce the cross-domain shift and provide a better alignment. We conduct a comprehensive empirical study on PointDA-10 and GraspNetPC-10 and show that our method achieves state-of the-art performance on GraspNetPC-10 (with approx. 4-12% margin) and best average performance on PointDA-10. Our ablation studies and decision boundary analysis also validate the significance of our contrastive learning module and OT alignment.

We propose an end-to-end inverse rendering pipeline called SupeRVol that allows us to recover 3D shape and material parameters from a set of color images in a super-resolution manner. To this end, we represent both the bidirectional reflectance distribution function (BRDF) and the signed distance function (SDF) by multi-layer perceptrons. In order to obtain both the surface shape and its reflectance properties, we revert to a differentiable volume renderer with a physically based illumination model that allows us to decouple reflectance and lighting. This physical model takes into account the effect of the camera's point spread function thereby enabling a reconstruction of shape and material in a super-resolution quality. Experimental validation confirms that SupeRVol achieves state of the art performance in terms of inverse rendering quality. It generates reconstructions that are sharper than the individual input images, making this method ideally suited for 3D modeling from low-resolution imagery.

We propose a discrete latent distribution for Generative Adversarial Networks (GANs). Instead of drawing latent vectors from a continuous prior, we sample from a finite set of learnable latents. However, a direct parametrization of such a distribution leads to an intractable linear increase in memory in order to ensure sufficient sample diversity. We address this key issue by taking inspiration from the encoding of information in biological organisms. Instead of learning a separate latent vector for each sample, we split the latent space into a set of genes. For each gene, we train a small bank of gene variants. Thus, by independently sampling a variant for each gene and combining them into the final latent vector, our approach can represent a vast number of unique latent samples from a compact set of learnable parameters. Interestingly, our gene-inspired latent encoding allows for new and intuitive approaches to latent-space exploration, enabling conditional sampling from our unconditionally trained model. Our approach preserves state-of-the-art photo-realism while achieving better disentanglement than the widely-used StyleMapping network.

Can a text-to-image diffusion model be used as a training objective for adapting a GAN generator to another domain? In this paper, we show that the classifier-free guidance can be leveraged as a critic and enable generators to distill knowledge from large-scale text-to-image diffusion models. Generators can be efficiently shifted into new domains indicated by text prompts without access to ground truth samples from target domains. We demonstrate the effectiveness and controllability of our method through extensive experiments. Although not trained to minimize CLIP loss, our model achieves equally high CLIP scores and significantly lower FID than prior work on short prompts and outperforms the baseline qualitatively and quantitatively on long and complicated prompts. To our best knowledge, the proposed method is the first attempt at incorporating large-scale pre-trained diffusion models and distillation sampling for text-driven image generator domain adaptation and gives a quality previously beyond possible. Moreover, we extend our work to 3D-aware style-based generators and DreamBooth guidance.

AR/VR applications promise to provide people with a genuine feeling of mutual presence when communicating via their personalized avatars. While realistic avatars are essential in various social settings, the vast possibilities of a virtual world can also generate interest in using stylized avatars for other purposes. We introduce StyleAvatar, the first method for semantic stylization of animatable head avatars. StyleAvatar directly stylizes the avatar representation, rather than stylizing its renders. Specifically, given a model generating the avatar, StyleAvatar first disentangles geometry and texture manipulations, and then stylizes the avatar by fine-tuning a subset of the model's weights. Our method has multiple virtues, including the ability to describe styles using images or text, preserving the avatar's animatable capacity, providing control over identity preservation, and disentangling texture and geometry modifications. Experiments have shown that our approach consistently works across skin tones, challenging hair styles, extreme views, and diverse facial expressions.

High-Definition (HD) maps are essential for the safety of autonomous driving systems. While existing techniques employ camera images and onboard sensors to generate vectorized high-precision maps, they are constrained by their reliance on single-frame input. This approach limits their stability and performance in complex scenarios such as occlusions, largely due to the absence of temporal information. Moreover, their performance diminishes when applied to broader perception ranges. In this paper, we present StreamMapNet, a novel online mapping pipeline adept at long-sequence temporal modeling of videos. StreamMapNet employs multi-point attention and temporal information which empowers the construction of large-range local HD maps with high stability and further addresses the limitations of existing methods. Furthermore, we critically examine widely used online HD Map construction benchmark and datasets, Argoverse2 and nuScenes, revealing significant bias in the existing evaluation protocols. We propose to resplit the benchmarks according to geographical spans, promoting fair and precise evaluations. Experimental results validate that StreamMapNet significantly outperforms existing methods across all settings while maintaining an online inference speed of 14.2 FPS. Our code is available at https://github.com/yuantianyuan01/StreamMapNet.

Universal domain adaptation (UniDA) is the unsupervised domain adaptation with label shift. UniDA aims to classify unlabeled target samples into one of the "known" categories or into a single "unknown" category. Its main challenge lies in detecting private classes from both domains and performing alignment between the common classes. Current methods employ various techniques and loss functions to address these challenges. However, these methods commonly represent classifiers as point weight vectors, which are prone to overfitting by the source domain samples due to the lack of supervision from the target domain. Consequently, these classifiers struggle to separate target samples into known and unknown categories effectively. To address this, we introduce a novel framework called Stochastic Binary Network for Universal Domain Adaptation (STUN). STUN uses a Stochastic binary classifier for each class, whose weight is modeled as Gaussian distribution, enabling to sample an arbitrary number of classifiers while keeping the model size same as of two classifiers. Consistency between these sampled classifiers is used to derive the confidence scores for both source and target samples, which facilitates the alignment of common classes using weighted adversarial learning. Finally, we use deep discriminative clustering to formulate a loss function for solving the problem of fragmented feature distributions in the target domain. Extensive ablation studies and state-of-the-art results across three standard benchmark datasets show the efficacy of our framework.

Real-time Stereo Matching is a cornerstone task for Extended Reality (XR) applications, such as 3D scene understanding, video pass-through, and mixed-reality games. Despite significant advancements, getting accurate depth information in real time on a low-power mobile device remains a challenge. One of the main difficulties is the lack of high-quality indoor video stereo data captured by head-mounted VR or AR glasses. To address this, we introduce a novel video stereo synthetic dataset that comprises photorealistic renderings of various indoor scenes and realistic camera motion captured by a moving VR/AR head-mounted display (HMD). Our newly proposed dataset enables one to develop a novel framework for continuous video-rate stereo matching. As another contribution, we also propose a new video-based stereo matching approach tailored for XR applications, which achieves real-time inference at an impressive 134fps on a standard desktop computer, or 30fps on a battery-powered HMD. Our key insight is that disparity and contextual information are highly correlated and redundant between consecutive stereo frames. By unrolling an iterative cost aggregation in time (i.e. in temporal dimension), we are able to distribute and reuse the aggregated features over time. This leads to a substantial reduction in computation without sacrificing accuracy. We conducted extensive evaluations and demonstrated that our method achieves superior performance compared to the current state-of-the-art, making it a strong contender for real-time stereo matching in VR/AR applications.

Despite of exciting advances in image-based rendering and novel view synthesis, it is still challenging to achieve high-resolution results that can reach production-level quality when applying such methods to the task of stereo conversion. At the same time, only very few dedicated stereo conversion approaches exist, which also fall short in terms of the required quality. Hence, in this paper, we present a novel method for high-resolution 2D-to-3D conversion. It is fully differentiable in all of its stages and performs disparity-informed warping, consistent foreground-background compositing, and background-aware inpainting. To enable temporal consistency in the resulting video, we propose a strategy to integrate information from additional video frames. Extensive ablation studies validate our design choices, leading to a fully automatic model that outperforms existing approaches by a large margin (49-70% LPIPS error reduction). Finally, inspired from current practices in manual stereo conversion, we introduce optional interactive tools into our model, which allow to steer the conversion process and make it significantly more applicable for 3D film production.