WACV 2024 Papers

Deep learning-based solutions for semantic segmentation suffer from significant performance degradation when tested on data with different characteristics than what was used during the training. Adapting the models using annotated data from the new domain is not always practical. Unsupervised Domain Adaptation (UDA) approaches are crucial in deploying these models in the actual operating conditions. Recent state-of-the-art (SOTA) UDA methods employ a teacher-student self-training approach, where a teacher model is used to generate pseudo-labels for the new data which in turn guide the training process of the student model. Though this approach has seen a lot of success, it suffers from the issue of noisy pseudo-labels being propagated in the training process. To address this issue, we propose an auxiliary pseudo-label refinement network (PRN) for online refining of the pseudo labels and also localizing the pixels whose predicted labels are likely to be noisy. Being able to improve the quality of pseudo labels and select highly reliable ones, PRN helps self-training of segmentation models to be robust against pseudo label noise propagation during different stages of adaptation. We evaluate our approach on benchmark datasets with three different domain shifts, and our approach consistently performs significantly better than the previous state-of-the-art methods.

Text-to-Image (T2I) synthesis is a challenging task requiring modelling both textual and image domains and their relationship. The substantial improvement in image quality achieved by recent works has paved the way for numerous applications such as language-aided image editing, computer-aided design, text-based image retrieval, and training data augmentation. In this work, we ask a simple question: Along with realistic images, can we obtain any useful by-product (e.g., foreground / background or multi-class segmentation masks, detection labels) in an unsupervised way that will also benefit other computer vision tasks and applications?. In an attempt to answer this question, we explore generating realistic images and their corresponding foreground / background segmentation masks from the given text. To achieve this, we experiment the concept of co-segmentation along with GAN. Specifically, a novel GAN architecture called Co-Segmentation Inspired GAN (COS-GAN) is proposed that generates two or more images simultaneously from different noise vectors and utilises a spatial co-attention mechanism between the image features to produce realistic segmentation masks for each of the generated images. The advantages of such an architecture are two-fold: 1) The generated segmentation masks can be used to focus on foreground and background exclusively to improve the quality of generated images, and 2) the segmentation masks can be used as a training target for other tasks, such as object localisation and segmentation. Extensive experiments conducted on CUB, Oxford-102, and COCO datasets show that COS-GAN is able to improve visual quality and generate reliable foreground / background masks for the generated images.

Most existing 3D human pose estimation work rely heavily on the powerful memory capability of networks to obtain suitable 2D-3D mappings from the training data. Few works have studied the modeling of human posture deformation in motion. In this paper, we propose a new modeling method for human pose deformations and design an accompanying diffusion-based motion prior. Inspired by the field of non-rigid structure-from-motion, we divide the task of reconstructing 3D human skeletons in motion into the estimation of a 3D reference skeleton, and a frame-by-frame skeleton deformation. A mixed spatial-temporal NRSfMformer is used to simultaneously estimate the 3D reference skeleton and the skeleton deformation of each frame from 2D observations sequence, and then sum them up to obtain the pose of each frame. Subsequently, a loss term based on the diffusion model is used to ensure that the pipeline learns the correct prior motion knowledge. Finally, we have evaluated our proposed method on mainstream datasets and obtained superior results outperforming the state-of-the-art.

Since distribution shifts are likely to occur during test-time and can drastically decrease the model's performance, online test-time adaptation (TTA) continues to update the model after deployment, leveraging the current test data. Clearly, a method proposed for online TTA has to perform well for all kinds of environmental conditions. By introducing the variable factors domain non-stationarity and temporal correlation, we first unfold all practically relevant settings and define the entity as universal TTA. We want to highlight that this is the first work that covers such a broad spectrum, which is indispensable for the use in practice. To tackle the problem of universal TTA, we identify and highlight several challenges a self-training based method has to deal with: 1) model bias and the occurrence of trivial solutions when performing entropy minimization on varying sequence lengths with and without multiple domain shifts, 2) loss of generalization which exacerbates the adaptation to multiple domain shifts and the occurrence of catastrophic forgetting, and 3) performance degradation due to shifts in class prior. To prevent the model from becoming biased, we leverage a dataset and model-agnostic certainty and diversity weighting. In order to maintain generalization and prevent catastrophic forgetting, we propose to continually weight-average the source and adapted model. To compensate for disparities in the class prior during test-time, we propose an adaptive prior correction scheme that reweights the model's predictions. We evaluate our approach, named ROID, on a wide range of settings, datasets, and models, setting new standards in the field of universal TTA. Code is available at: https://github.com/mariodoebler/test-time-adaptation

In this paper, we introduce a realistic and challenging domain adaptation problem called Universal Semi-supervised Model Adaptation (USMA), which i) requires only a pre-trained source model, ii) allows the source and target domain to have different label sets, i.e., they share a common label set and hold their own private label set, and iii) requires only a few labeled samples in each class of the target domain. To address USMA, we propose a collaborative consistency training framework that regularizes the prediction consistency between two models, i.e., a pre-trained source model and its variant pre-trained with target data only, and combines their complementary strengths to learn a more powerful model. The rationale of our framework stems from the observation that the source model performs better on common categories than the target-only model, while on target-private categories, the target-only model performs better. We also propose a two-perspective, i.e., sample-wise and class-wise, consistency regularization to improve the training. Experimental results demonstrate the effectiveness of our method on several benchmark datasets.

Given multiple videos of the same task, procedure learning addresses identifying the key-steps and determining their order to perform the task. For this purpose, existing approaches use the signal generated from a pair of videos. This makes key-steps discovery challenging as the algorithms lack inter-videos perspective. Instead, we propose an unsupervised Graph-based Procedure Learning (GPL) framework. GPL consists of the novel UnityGraph that represents all the videos of a task as a graph to obtain both intra-video and inter-videos context. Further, to obtain similar embeddings for the same key-steps, the embeddings of UnityGraph are updated in an unsupervised manner using the Node2Vec algorithm. Finally, to identify the key-steps, we cluster the embeddings using KMeans. We test GPL on benchmark ProceL, CrossTask, and EgoProceL datasets and achieve an average improvement of 2% on third-person datasets and 3.6% on EgoProceL over the state-of-the-art.

Text-to-image models suffer from various safety issues that may limit their suitability for deployment. Previous methods have separately addressed individual issues of bias, copyright, and offensive content in text-to-image models. However, in the real world, all of these issues appear simultaneously in the same model. We present a method that tackles all issues with a single approach. Our method, Unified Concept Editing (UCE), edits the model without training using a closed-form solution, and scales seamlessly to concurrent edits on text-conditional diffusion models. We demonstrate scalable simultaneous debiasing, style erasure, and content moderation by editing text-to-image projections, and we present extensive experiments demonstrating improved efficacy and scalability over prior work.

In recent years, there has been a growing trend of incorporating hyperbolic geometry methods into computer vision. While these methods have achieved state-of-the-art performance on various metric learning tasks using hyperbolic distance measurements, the underlying theoretical analysis supporting this superior performance remains under-exploited. In this study, we investigate the effects of integrating hyperbolic space into metric learning, particularly when training with contrastive loss. We identify a need for a comprehensive comparison between Euclidean and hyperbolic spaces regarding the temperature effect in the contrastive loss within the existing literature. To address this gap, we conduct an extensive investigation to benchmark the results of Vision Transformers (ViTs) using a hybrid objective function that combines loss from Euclidean and hyperbolic spaces. Additionally, we provide a theoretical analysis of the observed performance improvement. We also reveal that hyperbolic metric learning is highly related to hard negative sampling, providing insights for future work. This work will provide valuable data points and experience in understanding hyperbolic image embeddings. To shed more light on problem-solving and encourage further investigation into our approach, our code is available online.

Understanding dark scenes based on multi-modal image data is challenging, as both the visible and auxiliary modalities provide limited semantic information for the task. Previous methods focus on fusing the two modalities but neglect the correlations among semantic classes when minimizing losses to align pixels with labels, resulting in inaccurate class predictions. To address these issues, we introduce a supervised multi-modal contrastive learning approach to increase the semantic discriminability of the learned multi-modal feature spaces by jointly performing cross-modal and intra-modal contrast under the supervision of the class correlations. The cross-modal contrast encourages same-class embeddings from across the two modalities to be closer and pushes different-class ones apart. The intra-modal contrast forces same-class or different-class embeddings within each modality to be together or apart. We validate our approach on a variety of tasks that cover diverse light conditions and image modalities. Experiments show that our approach can effectively enhance dark scene understanding based on multi-modal images with limited semantics by shaping semantic-discriminative feature spaces. Comparisons with previous methods demonstrate our state-of-the-art performance. Code and pretrained models are available at https://github.com/palmdong/SMMCL.

State-of-the-art deep neural networks have been shown to be extremely powerful in a variety of perceptual tasks like semantic segmentation. However, these networks are vulnerable to adversarial perturbations of the input which are imperceptible for humans but lead to incorrect predictions. Treating image segmentation as a sum of pixel-wise classifications, adversarial attacks developed for classification models were shown to be applicable to segmentation models as well. In this work, we present simple uncertainty-based weighting schemes for the loss functions of such attacks that (i) put higher weights on pixel classifications which can more easily perturbed and (ii) zero-out the pixel-wise losses corresponding to those pixels that are already confidently misclassified. The weighting schemes can be easily integrated into the loss function of a range of well-known adversarial attackers with minimal additional computational overhead, but lead to significant improved perturbation performance, as we demonstrate in our empirical analysis on several datasets and models.

Instance segmentation has witnessed promising advancements through deep neural network-based algorithms. However, these models often exhibit incorrect predictions with unwarranted confidence levels. Consequently, evaluating prediction uncertainty becomes critical for informed decision-making. Existing methods primarily focus on quantifying uncertainty in classification or regression tasks, lacking emphasis on instance segmentation. Our research addresses the challenge of estimating spatial certainty associated with the location of instances with star-convex shapes. Two distinct clustering approaches are evaluated which compute spatial and fractional certainty per instance employing samples by the Monte-Carlo Dropout or Deep Ensemble technique. Our study demonstrates that combining spatial and fractional certainty scores yields improved calibrated estimation over individual certainty scores. Notably, our experimental results show that the Deep Ensemble technique alongside our novel radial clustering approach proves to be an effective strategy. Our findings emphasize the significance of evaluating the calibration of estimated certainties for model reliability and decision-making.

To reduce computation in deep neural network inference, a promising approach is to design a network with multiple internal classifiers (ICs) and adaptively select an execution path based on the complexity of a given input. However, quantizing an input-adaptive network, a must-do task for network deployment on edge devices, is a non-trivial task due to jointly allocating its computation budget along with network layers and IC locations. In this paper, we propose Unified Sample-wise Dynamic Network (USDN) with a mixed-precision and early-exit framework that obtains both the optimal location of ICs and layer-wise bit configurations under a given computation budget. The proposed USDN comprises multiple groups of layers, with each group representing a varying degree of complexity for input samples. Experimental results demonstrate that our approach reduces computational cost of the previous work by 12.78% while achieving higher accuracy on ImageNet dataset.

In this paper, we introduce UOW-Vessel, a benchmark dataset of high-resolution optical satellite images for vessel detection and segmentation. Our dataset consists of 3,500 images, collected from 14 countries across 4 continents. With a total of 35,598 instances in 10 vessel categories, UOW-Vessel is to date the largest satellite image dataset for vessel recognition. Furthermore, compared to the existing public datasets that only provide bounding box ground-truth, our new dataset offers more accurate polygon annotations of vessel objects. This dataset is expected to support instance segmentation-based approaches, which is a less investigated area in vessel surveillance. We also report extensive evaluations of the recent algorithms for instance segmentation on the new benchmark dataset.

We propose an efficient framework for 3D human pose and shape estimation from a video, named Uncertainty-Guided SpatioTemporal Transformer (UNSPAT). Unlike previous video-based methods that consider temporal relationships with global average pooled features, our approach incorporates both spatial and temporal dimensions without compromising spatial information. We address the excessive complexity of spatiotemporal attention through two modules: Spatial Alignment Module (SAM) and Space2Batch. The modules align input features and compute temporal attention at every spatial position in a batch-wise manner. Furthermore, our uncertainty-guided attention re-weighting module improves performance by diminishing the impact of artifacts. We demonstrate the effectiveness of the UNSPAT on widely used benchmark datasets and achieve state-of-the-art performance. Our method is robust to challenging scenes, such as occlusion, and cluttered backgrounds, showing its potential for real-world applications.

Recent image restoration methods can be broadly categorized into two classes: (1) regression methods that recover the rough structure of the original image without synthesizing high-frequency details and (2) generative methods that synthesize perceptually-realistic high-frequency details even though the resulting image deviates from the original structure of the input. While both directions have been extensively studied in isolation, merging their benefits with a single framework has been rarely studied. In this paper, we propose UGPNet, a universal image restoration framework that can effectively achieve the benefits of both approaches by simply adopting a pair of an existing regression model and a generative model. UGPNet first restores the image structure of a degraded input using a regression model and synthesizes a perceptually-realistic image with a generative model on top of the regressed output. UGPNet then combines the regressed output and the synthesized output, resulting in a final result that faithfully reconstructs the structure of the original image in addition to perceptually-realistic textures. Our extensive experiments on deblurring, denoising, and super-resolution demonstrate that UGPNet can successfully exploit both regression and generative methods for high-fidelity image restoration.

Contemporary point cloud segmentation approaches largely rely on richly annotated 3D training data. However, it is both time-consuming and challenging to obtain consistently accurate annotations for such 3D scene data. Moreover, there is still a lack of investigation into fully unsupervised scene segmentation for point clouds, especially for holistic 3D scenes. This paper presents U3DS3, as a step towards completely unsupervised point cloud segmentation for any holistic 3D scenes. To achieve this, U3DS3 leverages a generalized unsupervised segmentation method for both object and background across both indoor and outdoor static 3D point clouds with no requirement for model pre-training, by leveraging only the inherent information of the point cloud to achieve full 3D scene segmentation. The initial step of our proposed approach involves generating superpoints based on the geometric characteristics of each scene. Subsequently, it undergoes a learning process through a spatial clustering-based methodology, followed by iterative training using pseudo-labels generated in accordance with the cluster centroids. Moreover, by leveraging the invariance and equivariance of the volumetric representations, we apply the geometric transformation on voxelized features to provide two sets of descriptors for robust representation learning. Finally, our evaluation provides state-of-the-art results on the ScanNet and SemanticKITTI, and competitive results on the S3DIS, benchmark datasets.

Current state-of-the-art object proposal networks are trained with a closed-world assumption, meaning they learn to only detect objects of the training classes. These models fail to provide high recall in open-world environments where important novel objects may be encountered. While a handful of recent works attempt to tackle this problem, they fail to consider that the optimal behavior of a proposal network can vary significantly depending on the data and application. Our goal is to provide a flexible proposal solution that can be easily tuned to suit a variety of open-world settings. To this end, we design a Tunable Hybrid Proposal Network (THPN) that leverages an adjustable hybrid architecture, a novel self-training procedure, and dynamic loss components to optimize the tradeoff between known and unknown object detection performance. To thoroughly evaluate our method, we devise several new challenges which invoke varying degrees of label bias by altering known class diversity and label count. We find that in every task, THPN easily outperforms existing baselines (e.g., RPN, OLN). Our method is also highly data efficient, surpassing baseline recall with a fraction of the labeled data.

Spatiotemporal predictive learning offers a self-supervised learning paradigm that enables models to learn both spatial and temporal patterns by predicting future sequences based on historical sequences. Mainstream methods are dominated by recurrent units, yet they are limited by their lack of parallelization and often underperform in real-world scenarios. To improve prediction quality while maintaining computational efficiency, we propose an innovative triplet attention transformer designed to capture both inter-frame dynamics and intra-frame static features. Specifically, the model incorporates the Triplet Attention Module (TAM), which replaces traditional recurrent units by exploring self-attention mechanisms in temporal, spatial, and channel dimensions. In this configuration: (i) temporal tokens contain abstract representations of inter-frame, facilitating the capture of inherent temporal dependencies; (ii) spatial and channel attention combine to refine the intra-frame representation by performing fine-grained interactions across spatial and channel dimensions. Alternating temporal, spatial, and channel-level attention allows our approach to learn more complex short- and long-range spatiotemporal dependencies. Extensive experiments demonstrate performance surpassing existing recurrent-based and recurrent-free methods, achieving state-of-the-art under multi-scenario examination including moving object trajectory prediction, traffic flow prediction, driving scene prediction, and human motion capture.

Recent progress in NeRF-based GANs has introduced a number of approaches for high-resolution and high-fidelity generative modeling of human heads with a possibility for novel view rendering. At the same time, one must solve an inverse problem to be able to re-render or modify an existing image or video. Despite the success of universal optimization-based methods for 2D GAN inversion, those applied to 3D GANs may fail to extrapolate the result onto the novel view, whereas optimization-based 3D GAN inversion methods are time-consuming and can require at least several minutes per image. Fast encoder-based techniques, such as those developed for StyleGAN, may also be less appealing due to the lack of identity preservation. Our work introduces a fast technique that bridges the gap between the two approaches by directly utilizing the tri-plane representation presented for the EG3D generative model. In particular, we build upon a feed-forward convolutional encoder for the latent code and extend it with a fully-convolutional predictor of tri-plane numerical offsets. The renderings are similar in quality to the ones produced by optimization-based techniques and outperform the ones by encoder-based methods. As we empirically prove, this is a consequence of directly operating in the tri-plane space, not in the GAN parameter space, while making use of an encoder-based trainable approach. Finally, we demonstrate significantly more correct embedding of a face image in 3D than for all the baselines, further strengthened by a probably symmetric prior enabled during training.