ICCV 2023 Papers

Learning novel concepts, remembering previous knowledge, and adapting it to future tasks occur simultaneously throughout a human's lifetime. To model such comprehensive abilities, continual zero-shot learning (CZSL) has recently been introduced. However, most existing methods overused the unseen semantic information that may not be continually accessible in realistic settings. In this paper, we address the challenge of continual zero-shot learning where unseen information is not provided during training, by leveraging generative modeling. The heart of the generative-based methods is to learn quality representations from seen classes to improve the generative understanding of the unseen visual space. Motivated by this, we introduce generalization-bound tools and provide the first theoretical explanation for the benefits of generative modeling to CZSL tasks. Guided by the theoretical analysis, we then propose our learning algorithm that employs a novel semantically guided Generative Random Walk (GRW) loss. The GRW loss augments the training by continually encouraging the model to generate realistic and characterized samples to represent the unseen space. Our algorithm achieves state-of-the-art performance on AWA1, AWA2, CUB, and SUN datasets, surpassing existing CZSL methods by 3-7%. The code is available here https://github.com/wx-zhang/IGCZSL

Deep learning empowers the mainstream medical image segmentation methods. Nevertheless, current deep segmentation approaches are not capable of efficiently and effectively adapting and updating the trained models when new segmentation classes are incrementally added. In the real clinical environment, it can be preferred that segmentation models could be dynamically extended to segment new organs/tumors without the (re-)access to previous training datasets due to obstacles of patient privacy and data storage. This process can be viewed as a continual semantic segmentation (CSS) problem, being understudied for multi-organ segmentation. In this work, we propose a new architectural CSS learning framework to learn a single deep segmentation model for segmenting a total of 143 whole-body organs. Using the encoder/decoder network structure, we demonstrate that a continually trained then frozen encoder coupled with incrementally-added decoders can extract sufficiently representative image features for new classes to be subsequently and validly segmented, while avoiding the catastrophic forgetting in CSS. To maintain a single network model complexity, each decoder is progressively pruned using neural architecture search and teacher-student based knowledge distillation. Finally, we propose a body-part and anomaly-aware output merging module to combine organ predictions originating from different decoders and incorporate both healthy and pathological organs appearing in different datasets. Trained and validated on 3D CT scans of 2500+ patients from four datasets, our single network can segment a total of 143 whole-body organs with very high accuracy, closely reaching the upper bound performance level by training four separate segmentation models (i.e., one model per dataset/task).

Co-speech gestures are a key channel of human communication, making them important for personalized chat agents to generate. In the past, gesture generation models assumed that data for each speaker is available all at once, and in large amounts. However in practical scenarios, speaker data comes sequentially and in small amounts as the agent personalizes with more speakers, akin to a continual learning paradigm. While more recent works have shown progress in adapting to low-resource data, they catastrophically forget the gesture styles of initial speakers they were trained on. Also, prior generative continual learning works are not multimodal, making this space less studied. In this paper, we explore this new paradigm and propose C-DiffGAN: an approach that continually learns new speaker gesture styles with only a few minutes of per-speaker data, while retaining previously learnt styles. Inspired by prior continual learning works, C-DiffGAN encourages knowledge retention by 1) generating reminiscences of previous low-resource speaker data, then 2) crossmodally aligning to them to mitigate catastrophic forgetting. We quantitatively demonstrate improved performance and reduced forgetting over strong baselines through standard continual learning measures, reinforced by a qualitative user study that shows that our method produces more natural, style-preserving gestures. Code and videos can be found at https://chahuja.com/cdiffgan

Accomplishing household tasks such as 'bringing a cup of water' requires to plan step-by-step actions by maintaining the knowledge about the spatial arrangement of objects and consequences of previous actions. Perception models of current embodied AI agents, however, often make mistakes due to lack of such knowledge but rely on imperfect learning of imitating agents or an algorithmic planner without the knowledge about the changed environment by the previous actions. To address the issue, we propose the CPEM (Context-aware Planner and Environment-aware Memory) embodied agent to incorporate the contextual information of previous actions for planning and maintaining spatial arrangement of objects with their states (e.g., if an object has been already moved or not) in the environment to the perception model for improving both visual navigation and object interactions. We observe that the proposed model achieves state-of-the-art task success performance in various metrics using a challenging interactive instruction following benchmark both in seen and unseen environments by large margins (up to +10.70% in unseen env.).

Recently, GAN has successfully contributed to making single-image super-resolution (SISR) methods produce more realistic images. However, natural images have complex distribution in the real world, and a single classifier in the discriminator may not have enough capacity to classify real and fake samples, making the preceding SR network generate unpleasing noise and artifacts. To solve the problem, we propose a novel content-aware local GAN framework, CAL-GAN, which processes a large and complicated distribution of real-world images by dividing them into smaller subsets based on similar contents. Our mixture of classifiers (MoC) design allocates different super-resolved patches to corresponding expert classifiers. Additionally, we introduce novel routing and orthogonality loss terms so that different classifiers can handle various contents and learn separable features. By feeding similar distributions into the corresponding specialized classifiers, CAL-GAN enhances the representation power of existing super-resolution models, achieving state-of-the-art perceptual performance on standard benchmarks and real-world images without modifying the generator-side architecture.

Remote photoplethysmography (rPPG) is a promising research area involving non-invasive monitoring of vital signs using cameras. While several supervised methods have been proposed, recent research has focused on contrastive-based self-supervised methods. However, these methods often collapse to learning irrelevant periodicities when dealing with interferences such as head motions, facial dynamics, and video compression. To address this limitation, firstly, we enhance the current self-supervised learning by introducing more reliable and explicit contrastive constraints. Secondly, we propose an innovative learning strategy that seamlessly integrates self-supervised constraints with pseudo-supervisory signals derived from traditional unsupervised methods. This is followed by a co-rectification technique designed to mitigate the adverse effects of noisy pseudo-labels. Experimental results demonstrate the superiority of our methodology over representative models when applied to small, high-quality datasets such as PURE and UBFC-rPPG. Importantly, on large-scale challenging datasets such as VIPL-HR and V4V, our method, with zero annotation cost, not only significantly surpasses prevailing self-supervised techniques but also showcases remarkable alignment with state-of-the-art supervised methods.

This paper presents a novel object-centric contact representation ContactGen for hand-object interaction. The ContactGen comprises 3 components: a contact map indicates the contact location, a part map represents the contact hand part, and a direction map tells the contact direction within each part. Given an input object, we propose a conditional generative model to predict ContactGen and adopt model-based optimization to predict diverse and geometrically feasible grasps. Experimental results demonstrate our method can generate high-fidelity and diverse human grasps for various objects.

Learning-based multi-view stereo (MVS) methods deal with predicting accurate depth maps to achieve an accurate and complete 3D representation. Despite the excellent performance, existing methods ignore the fact that a suitable depth geometry is also critical in MVS. In this paper, we demonstrate that different depth geometries have significant performance gaps, even using the same depth prediction error. Therefore, we introduce an ideal depth geometry composed of Saddle-Shaped Cells, whose predicted depth map oscillates upward and downward around the ground-truth surface, rather than maintaining a continuous and smooth depth plane. To achieve it, we develop a coarse-to-fine framework called Dual-MVSNet (DMVSNet), which can produce an oscillating depth plane. Technically, we predict two depth values for each pixel (Dual-Depth) and propose a novel loss function and a checkerboard-shaped selecting strategy to constrain the predicted depth geometry. Compared to existing methods, DMVSNet achieves a high rank on the DTU benchmark and obtains the top performance on challenging scenes of Tanks and Temples, demonstrating its strong performance and generalization ability. Our method also points to a new research direction for considering depth geometry in MVS.

Transparent objects are commonly seen in indoor scenes but are hard to estimate. Currently, commercial depth cameras face difficulties in estimating the depth of transparent objects due to the light reflection and refraction on their surface. As a result, they tend to make a noisy and incorrect depth value for transparent objects. These incorrect depth data make the traditional RGB-D SLAM method fails in reconstructing the scenes that contain transparent objects. An exact depth value of the transparent object is required to restore in advance and it is essential that the depth value of the transparent object must keep consistent in different views, or the reconstruction result will be distorted. Previous depth prediction methods of transparent objects can restore these missing depth values but none of them can provide a good result in reconstruction due to the inconsistency prediction. In this work, we propose a real-time reconstruction method using a novel stereo-based depth prediction network to keep the consistency of depth prediction in a sequence of images. Because there is no video dataset about transparent objects currently to train our model, we construct a synthetic RGB-D video dataset with different transparent objects. Moreover, to test generalization capability, we capture video from real scenes using the RealSense D435i RGB-D camera. We compare the metrics on our dataset and SLAM reconstruction results in both synthetic scenes and real scenes with the previous methods. Experiments show our significant improvement in accuracy on depth prediction and scene reconstruction.

Unsupervised domain adaptation aims to transfer knowledge from a fully-labeled source domain to an unlabeled target domain. However, in real-world scenarios, providing abundant labeled data even in the source domain can be infeasible due to the difficulty and high expense of annotation. To address this issue, recent works consider the Few-shot Unsupervised Domain Adaptation (FUDA) where only a few source samples are labeled, and conduct knowledge transfer via self-supervised learning methods. Yet existing methods generally overlook that the sparse label setting hinders learning reliable source knowledge for transfer. Additionally, the learning difficulty difference in target samples is different but ignored, leaving hard target samples poorly classified. To tackle both deficiencies, in this paper, we propose a novel Confidence-based Visual Dispersal Transfer learning method (C-VisDiT) for FUDA. Specifically, C-VisDiT consists of a cross-domain visual dispersal strategy that transfers only high-confidence source knowledge for model adaptation and an intra-domain visual dispersal strategy that guides the learning of hard target samples with easy ones. We conduct extensive experiments on Office-31, Office-Home, VisDA-C, and DomainNet benchmark datasets and the results demonstrate that the proposed C-VisDiT significantly outperforms state-of-the-art FUDA methods. Our code is available at https://github.com/Bostoncake/C-VisDiT.

Visual grounding aims at localizing the target object in image which is most related to the given free-form natural language query. As labeling the position of target object is labor-intensive, the weakly supervised methods, where only image-sentence annotations are required during model training have recently received increasing attention. Most of the existing weakly-supervised methods first generate region proposals via pre-trained object detectors and then employ either cross-modal similarity score or reconstruction loss as the criteria to select proposal from them. However, due to the cross-modal heterogeneous gap, these method often suffer from high confidence spurious association and model prone to error propagation. In this paper, we propose Confidence-aware Pseudo-label Learning (CPL) to overcome the above limitations. Specifically, we first adopt both the uni-modal and cross-modal pre-trained models and propose conditional prompt engineering to automatically generate multiple `descriptive, realistic and diverse' pseudo language queries for each region proposal, and then establish reliable cross-modal association for model training based on the uni-modal similarity score (between pseudo and real text queries). Secondly, we propose a confidence-aware pseudo label verification module which reduces the amount of noise encountered in the training process and the risk of error propagation. Experiments on five widely used datasets validate the efficacy of our proposed components and demonstrate state-of-the-art performance.

Many studies in vision tasks have aimed to create effective embedding spaces for single-label object prediction within an image. However, in reality, most objects possess multiple specific attributes, such as shape, color, and length, with each attribute composed of various classes. To apply models in real-world scenarios, it is essential to be able to distinguish between the granular components of an object. Conventional approaches to embedding multiple specific attributes into a single network often result in entanglement, where fine-grained features of each attribute cannot be identified separately. To address this problem, we propose a Conditional Cross-Attention Network that induces disentangled multi-space embeddings for various specific attributes with only a single backbone. Firstly, we employ a cross-attention mechanism to fuse and switch the information of conditions (specific attributes), and we demonstrate its effectiveness through a diverse visualization example. Secondly, we leverage the vision transformer for the first time to a fine-grained image retrieval task and present a simple yet effective framework compared to existing methods. Unlike previous studies where performance varied depending on the benchmark dataset, our proposed method achieved consistent state-of-the-art performance on the FashionAI, DARN, DeepFashion, and Zappos50K benchmark datasets.

In this paper, we address the problem of conditional scene decoration for 360deg images. Our method takes a 360deg background photograph of an indoor scene and generates decorated images of the same scene in the panorama view. To do this, we develop a 360-aware object layout generator that learns latent object vectors in the 360deg view to enable a variety of furniture arrangements for an input 360deg background image. We use this object layout to condition a generative adversarial network to synthesize images of an input scene. To further reinforce the generation capability of our model, we develop a simple yet effective scene emptier that removes the generated furniture and produces an emptied scene for our model to learn a cyclic constraint. We train the model on the Structure3D dataset and show that our model can generate diverse decorations with controllable object layout. Our method achieves state-of-the-art performance on the Structure3D dataset and generalizes well to the Zillow indoor scene dataset. Our user study confirms the immersive experiences provided by the realistic image quality and furniture layout in our generation results. Our implementation is available at https://github.com/kcshum/neural_360_decoration.git.

Generative Adversarial Networks (GANs) can produce photo-realistic results using an unconditional image-generation pipeline. However, the images generated by GANs (e.g., StyleGAN) are entangled in feature spaces, which makes it difficult to interpret and control the contents of images. In this paper, we present an encoder-decoder model that decomposes the entangled GAN space into a conceptual and hierarchical latent space in a self-supervised manner. The outputs of 3D morphable face models are leveraged to independently control image synthesis parameters like pose, expression, and illumination. For this purpose, a novel latent space decomposition pipeline is introduced using transformer networks and generative models. Later, this new space is used to optimize a transformer-based GAN space controller for face editing. In this work, a StyleGAN2 model for faces is utilized. Since our method manipulates only GAN features, the photo-realism of StyleGAN2 is fully preserved. The results demonstrate that our method qualitatively and quantitatively outperforms baselines in terms of identity preservation and editing precision.

A multitude of prevalent pre-trained models mark a major milestone in the development of artificial intelligence, while fine-tuning has been a common practice that enables pre-trained models to figure prominently in a wide array of target datasets. Our empirical results reveal that off-the-shelf fine-tuning techniques are far from adequate to mitigate negative transfer caused by two types of underperforming features in a pre-trained model, including rare features and spuriously correlated features. Rooted in structural causal models of predictions after fine-tuning, we propose a Concept-wise fine-tuning (Concept-Tuning) approach which refines feature representations in the level of patches with each patch encoding a concept. Concept-Tuning minimizes the negative impacts of rare features and spuriously correlated features by (1) maximizing the mutual information between examples in the same category with regard to a slice of rare features (a patch) and (2) applying front-door adjustment via attention neural networks in channels and feature slices (patches). The proposed Concept-Tuning consistently and significantly (by up to 4.76%) improves prior state-of-the-art fine-tuning methods on eleven datasets, diverse pre-training strategies (supervised and self-supervised ones), various network architectures, and sample sizes in a target dataset.

Whole slide image (WSI) analysis has become increasingly important in the medical imaging community, enabling automated and objective diagnosis, prognosis, and therapeutic-response prediction. However, in clinical practice, the continuous progress of evolving WSI acquisition technology, the diversity of scanners, and different imaging protocols hamper the utility of WSI analysis models. In this paper, we propose the FIRST continual learning framework for WSI analysis, named ConSlide, to tackle the challenges of enormous image size, utilization of hierarchical structure, and catastrophic forgetting by progressive model updating on multiple sequential datasets. Our framework contains three key components. The Hierarchical Interaction Transformer (HIT) is proposed to model and utilize the hierarchical structural knowledge of WSI. The BreakupReorganize (BuRo) rehearsal method is developed for WSI data replay with efficient region storing buffer and WSI reorganizing operation. The asynchronous updating mechanism is devised to encourage the network to learn generic and specific knowledge respectively during the replay stage, based on a nested cross-scale similarity learning (CSSL) module. We evaluated the proposed ConSlide on four public WSI datasets from TCGA projects. It performs best over other state-of-the-art methods with a fair WSI-based continual learning setting and achieves a better trade-off of the overall performance and forgetting on previous tasks.

Neural image compression methods have seen increasingly strong performance in recent years. However, they suffer orders of magnitude higher computational complexity compared to traditional codecs, which hinders their real-world deployment. This paper takes a step forward in closing this gap in decoding complexity by adopting shallow or even linear decoding transforms. To compensate for the resulting drop in compression performance, we exploit the often asymmetrical computation budget between encoding and decoding, by adopting more powerful encoder networks and iterative encoding. We theoretically formalize the intuition behind, and our experimental results establish a new frontier in the trade-off between rate-distortion and decoding complexity for neural image compression. Specifically, we achieve rate-distortion performance competitive with the established mean-scale hyperprior architecture of Minnen et al. (2018) at less than 50K decoding FLOPs/pixel, reducing the baseline's overall decoding complexity by 80%, or over 90% for the synthesis transform alone. Our code can be found at https://github.com/mandt-lab/shallow-ntc.

Underlying many structured light systems, especially those based on laser scanning, is a simple vision task: tracking a light spot. To accomplish this, scanners use conventional CMOS sensors to capture, transmit, and process millions of pixel measurements. This approach, while capable of achieving high-fidelity 3D scans, is wasteful in terms of (often scarce) sensing and computational resources. We present a structured light system based on position sensing diodes (PSDs), an unconventional sensing modality that directly measures the centroid of the spatial distribution of incident light, thus enabling high-resolution 3D laser scanning with a minimal amount of sensor data. We develop theory and computational algorithms for PSD-based structured light under a variety of light transport effects. We demonstrate the benefits of the proposed techniques using a hardware prototype on several real-world scenes, including optically-challenging objects with long-range inter-reflections and scattering.

Backdoor attacks against deep learning have been widely studied. Various attack techniques have been proposed for different domains and paradigms, e.g., image, point cloud, natural language processing, transfer learning, etc. These works normally adopt the data poisoning strategy to embed the backdoor. They randomly select samples from the benign training set for poisoning, without considering the distinct contribution of each sample to the backdoor effectiveness, making the attack less optimal. A recent work (IJCAI-22) proposed to use the forgetting score to measure the importance of each poisoned sample and then filter out redundant data for effective backdoor training. However, this method is empirically designed without theoretical proofing. It is also very time-consuming as it needs to go through almost all the training stages for data selection. To address such limitations, we propose a novel confidence-based scoring methodology, which can efficiently measure the contribution of each poisoning sample based on the distance posteriors. We further introduce a greedy search algorithm to find the most informative samples for backdoor injection more promptly. Experimental evaluations on both 2D image and 3D point cloud classification tasks show that our approach can achieve comparable performance or even surpass the forgetting score-based searching method while requiring only several extra epochs' computation of a standard training process.