ICCV 2023 Papers

Video-language pre-training has advanced the performance of various downstream video-language tasks. However, most previous methods directly inherit or adapt typical image-language pre-training paradigms to video-language pre-training, thus not fully exploiting the unique characteristic of video, i.e., temporal. In this paper, we propose a Hierarchical Temporal-Aware video-language pre-training framework, HiTeA, with two novel pre-training tasks for yielding temporal-aware multi-modal representation with cross-modal fine-grained temporal moment information and temporal contextual relations between video-text multi-modal pairs. First, we propose a cross-modal moment exploration task to explore moments in videos by mining the paired texts, which results in detailed video moment representation. Then, based on the learned detailed moment representations, the inherent temporal contextual relations are captured by aligning video-text pairs as a whole in different time resolutions with multi-modal temporal relation exploration task. Furthermore, we introduce the shuffling test to evaluate the temporal reliance of datasets and video-language pre-training models. We achieve state-of-the-art results on 15 well-established video-language understanding and generation tasks, especially on temporal-oriented datasets (e.g., SSv2-Template and SSv2-Label) with 8.6% and 11.1% improvement respectively. HiTeA also demonstrates strong generalization ability when directly transferred to downstream tasks in a zero-shot manner.

Panoptic Scene Graph generation (PSG) is a recently proposed task in image scene understanding that aims to segment the image and extract triplets of subjects, objects and their relations to build a scene graph. This task is particularly challenging for two reasons. First, it suffers from a long-tail problem in its relation categories, making naive biased methods more inclined to high-frequency relations. Existing unbiased methods tackle the long-tail problem by data/loss rebalancing to favor low-frequency relations. Second, a subject-object pair can have two or more semantically overlapping relations. While existing methods favor one over the other, our proposed HiLo framework lets different network branches specialize on low and high frequency relations, enforce their consistency and fuse the results. To the best of our knowledge we are the first to propose an explicitly unbiased PSG method. In extensive experiments we show that our HiLo framework achieves state-of-the-art results on the PSG task. We also apply our method to the Scene Graph Generation task that predicts boxes instead of masks and see improvements over all baseline methods. Code is available at https://github.com/franciszzj/HiLo.

3D Morphable Models (3DMMs) demonstrate great potential for reconstructing faithful and animatable 3D facial surfaces from a single image. The facial surface is influenced by the coarse shape, as well as the static detail (e,g., person-specific appearance) and dynamic detail (e.g., expression-driven wrinkles). Previous work struggles to decouple the static and dynamic details through image-level supervision, leading to reconstructions that are not realistic. In this paper, we aim at high-fidelity 3D face reconstruction and propose HiFace to explicitly model the static and dynamic details. Specifically, the static detail is modeled as the linear combination of a displacement basis, while the dynamic detail is modeled as the linear interpolation of two displacement maps with polarized expressions. We exploit several loss functions to jointly learn the coarse shape and fine details with both synthetic and real-world datasets, which enable HiFace to reconstruct high-fidelity 3D shapes with animatable details. Extensive quantitative and qualitative experiments demonstrate that HiFace presents state-of-the-art reconstruction quality and faithfully recovers both the static and dynamic details.

Class-incremental learning (CIL) has achieved remarkable successes in learning new classes consecutively while overcoming catastrophic forgetting on old categories. However, most existing CIL methods unreasonably assume that all old categories have the same forgetting pace, and neglect negative influence of forgetting heterogeneity among different old classes on forgetting compensation. To surmount the above challenges, we develop a novel Heterogeneous Forgetting Compensation (HFC) model, which can resolve heterogeneous forgetting of easy-to-forget and hard-to-forget old categories from both representation and gradient aspects. Specifically, we design a task-semantic aggregation block to alleviate heterogeneous forgetting from representation aspect. It aggregates local category information within each task to learn task-shared global representations. Moreover, we develop two novel plug-and-play losses: a gradient-balanced forgetting compensation loss and a gradient-balanced relation distillation loss to alleviate forgetting from gradient aspect. They consider gradient-balanced compensation to rectify forgetting heterogeneity of old categories and heterogeneous relation consistency. Experiments on several representative datasets illustrate effectiveness of our HFC model. The code is available at https://github.com/JiahuaDong/HFC.

The existing one-stage multi-object tracking (MOT) algorithms have achieved satisfactory performance benefiting from a large amount of labeled data. However, acquiring plenty of laborious annotated frames is not practical in real applications. To reduce the cost of human annotations, we propose Heterogeneous Diversity driven Active Multi-Object Tracking (HD-AMOT), to infer the most informative frames for any MOT tracker by observing the heterogeneous cues of samples. HD-AMOT defines the diversified informative representation by encoding the geometric and semantic information, and formulates the frame inference strategy as a Markov decision process to learn an optimal sampling policy based on the designed informative representation. Specifically, HD-AMOT consists of a diversified informative representation module as well as an informative frame selection network. The former produces the signal characterizing the diversity and distribution of frames, and the latter receives the signal and conducts multi-frame cooperation to enable batch frame sampling. Extensive experiments conducted on the MOT15, MOT17, MOT20, and Dancetrack datasets demonstrate the efficacy and effectiveness of HD-AMOT. Experiments show that under 50% budget our HD-AMOT can achieve similar or even higher performance as fully-supervised learning.

We introduce an object-aware decoder for improving the performance of spatio-temporal representations on ego-centric videos. The key idea is to enhance object-awareness during training by tasking the model to predict hand positions, object positions, and the semantic label of the objects using paired captions when available. At inference time the model only requires RGB frames as inputs, and is able to track and ground objects (although it has not been trained explicitly for this). We demonstrate the performance of the object-aware representations learnt by our model, by: (i) evaluating it for strong transfer, i.e, through zero-shot testing, on a number of downstream video-text retrieval and classification benchmarks; and (ii) by evaluating its temporal and spatial (grounding) performance by fine-tuning for this task. In all cases the performance improves over the state of the art -- even for networks trained with far larger batch sizes. Overall, we show that the model can act as a drop-in replacement for an ego-centric video model, and improve performance.

We present a video decomposition method that facilitates layer-based editing of videos with spatiotemporally varying lighting and motion effects. Our neural model decomposes an input video into multiple layered representations, each comprising a 2D texture map, a mask for the original video, and a multiplicative residual characterizing the spatiotemporal variations in lighting conditions. A single edit on the texture maps can be propagated to the corresponding locations in the entire video frames while preserving other contents' consistencies. Our method efficiently learns the layer-based neural representations of a 1080p video in 25s per frame via coordinate hashing and allows real-time rendering of the edited result at 71 fps on a single GPU. Qualitatively, we run our method on various videos to show its effectiveness in generating high-quality editing effects. Quantitatively, we propose to adopt feature-tracking evaluation metrics for objectively assessing the consistency of video editing. Project page: https://lightbulb12294.github.io/hashing-nvd/

Glaucoma is the number one cause of irreversible blindness globally. A major challenge for accurate glaucoma detection and progression forecasting is the bottleneck of limited labeled patients with the state-of-the-art (SOTA) 3D retinal imaging data of optical coherence tomography (OCT). To address the data scarcity issue, this paper proposes two solutions. First, we develop a novel generalization-reinforced semi-supervised learning (SSL) model called pseudo supervisor to optimally utilize unlabeled data. Compared with SOTA models, the proposed pseudo supervisor optimizes the policy of predicting pseudo labels with unlabeled samples to improve empirical generalization. Our pseudo supervisor model is evaluated with two clinical tasks consisting of glaucoma detection and progression forecasting. The progression forecasting task is evaluated both unimodally and multimodally. Our pseudo supervisor model demonstrates superior performance than SOTA SSL comparison models. Moreover, our model also achieves the best results on the publicly available LAG fun- dus dataset. Second, we introduce the Harvard Glaucoma Detection and Progression (Harvard-GDP) Dataset, a multimodal multitask dataset that includes data from 1,000 patients with OCT imaging data, as well as labels for glaucoma detection and progression. This is the largest glaucoma detection dataset with 3D OCT imaging data and the first glaucoma progression forecasting dataset that is publicly available. Detailed sex and racial analysis are pro- vided, which can be used by interested researchers for fairness learning studies. Our released dataset is benchmarked with several SOTA supervised CNN and transformer deep learning models. The dataset and code are made publicly available via https://ophai.hms.harvard.edu/ datasets/harvard-gdp1000.

Diffusion-based models have achieved state-of-the-art performance on text-to-image synthesis tasks. However, one critical limitation of these models is the low fidelity of generated images with respect to the text description, such as missing objects, mismatched attributes, and mislocated objects. One key reason for such inconsistencies is the inaccurate cross-attention to text in both the spatial dimension, which controls at what pixel region an object should appear, and the temporal dimension, which controls how different levels of details are added through the denoising steps. In this paper, we propose a new text-to-image algorithm that adds explicit control over spatial-temporal cross-attention in diffusion models. We first utilize a layout predictor to predict the pixel regions for objects mentioned in the text. We then impose spatial attention control by combining the attention over the entire text description and that over the local description of the particular object in the corresponding pixel region of that object. The temporal attention control is further added by allowing the combination weights to change at each denoising step, and the combination weights are optimized to ensure high fidelity between the image and the text. Experiments show that our method generates images with higher fidelity compared to diffusion-model-based baselines without fine-tuning the diffusion model. Our code is publicly available.

Recently, methods for skeleton-based human activity recognition have been shown to be vulnerable to adversarial attacks. However, these attack methods require either the full knowledge of the victim (i.e. white-box attacks), access to training data (i.e. transfer-based attacks) or frequent model queries (i.e. black-box attacks). All their requirements are highly restrictive, raising the question of how detrimental the vulnerability is. In this paper, we show that the vulnerability indeed exists. To this end, we consider a new attack task: the attacker has no access to the victim model or the training data or labels, where we coin the term hard no-box attack. Specifically, we first learn a motion manifold where we define an adversarial loss to compute a new gradient for the attack, named skeleton-motion-informed (SMI) gradient. Our gradient contains information of the motion dynamics, which is different from existing gradient-based attack methods that compute the loss gradient assuming each dimension in the data is independent. The SMI gradient can augment many gradient-based attack methods, leading to a new family of no-box attack methods. Extensive evaluation and comparison show that our method imposes a real threat to existing classifiers. They also show that the SMI gradient improves the transferability and imperceptibility of adversarial samples in both no-box and transfer-based black-box settings.

While analyzing scanned documents, handwritten text can overlap with printed text. This overlap causes difficulties during the optical character recognition (OCR) and digitization process of documents, and subsequently, hurts downstream NLP tasks. Prior research either focuses solely on the binary classification of handwritten text or performs a three-class segmentation of the document, i.e., recognition of handwritten, printed, and background pixels. This approach results in the assignment of overlapping handwritten and printed pixels to only one of the classes, and thus, they are not accounted for in the other class. Thus, in this research, we develop novel approaches to address the challenges of handwritten and printed text segmentation. Our objective is to recover text from different classes in their entirety, especially enhancing the segmentation performance on overlapping sections. To support this task, we introduce a new dataset, SignaTR6K, collected from real legal documents, as well as a new model architecture for the handwritten and printed text segmentation task. Our best configuration outperforms prior work on two different datasets by 17.9% and 7.3% on IoU scores. The SignaTR6K dataset is accessible for download via the following link: https://forms.office.com/r/2a5RDg7cAY.

3D hand pose estimation is a critical task in various human-computer interaction applications. Numerous deep learning based estimation models in this domain have been actively explored. However, the existing models follows a non-recurrent scheme and thus require complex architectures or redundant parameters in order to achieve acceptable model capacity. To tackle this limitation, this paper proposes HandR2N2, a compact neural network that iteratively regresses the hand pose using a novel residual recurrent unit. The recurrent design allows recursive exploitation of partial layers to gradually optimize previously estimated joint locations. In addition, we exploit graph reasoning to capture kinematic dependencies between joints for better performance. Experimental results show that the proposed model significantly outperforms the existing methods on three hand pose benchmark datasets in terms of both accuracy and efficiency. Codes and pre-trained models are publicly available at https://github.com/cwc1260/HandR2N2.

Contrastive learning models based on Siamese structure have demonstrated remarkable performance in self-supervised learning. Such a success of contrastive learning relies on two conditions, including a sufficient number of positive pairs and adequate variations between them. If the conditions are not met, these frameworks will lack semantic contrast and be fragile on overfitting. To address these two issues, we propose Hallucinator that could efficiently generate additional positive samples for further contrast. The Hallucinator creates new data in the feature space, thus introducing nearly negligible computation. Moreover, we reduce the mutual information of hallucinated pairs and smooth them through non-linear operations. This process helps avoid over-confident contrastive learning models during the training and achieves more robust transformation-invariant feature embeddings. Remarkably, we empirically prove that the proposed Hallucinator generalizes well to various contrastive learning models, including MoCoV1&V2, SimCLR and SimSiam. Under the linear classification protocol, a stable accuracy gain is achieved, ranging from 0.3% to 3.0% on CIFAR10&100, Tiny ImageNet, STL-10 and ImageNet. The improvement is also observed in transferring pre-train encoders to the downstream tasks, including object detection and segmentation.

We propose a novel hairstyle transferred image synthesis method considering the underlying head geometry of two input images. In traditional GAN-based methods, transferring hairstyle from one image to the other often makes the synthesized result awkward due to differences in pose, shape, and size of heads. To resolve this, we utilize neural rendering by registering two input heads in the volumetric space to make a transferred hairstyle fit on the head of a target image. Because of the geometric nature of neural rendering, our method can render view varying images of synthesized results from a single transfer process without causing distortion from which extant hairstyle transfer methods built upon traditional GAN-based generators suffer. We verify that our method surpasses other baselines in view of preserving the identity and hairstyle of two input images when synthesizing a hairstyle transferred image rendered at any point of view.

Hair editing has made tremendous progress in recent years. Early hair editing methods use well-drawn sketches or masks to specify the editing conditions. Even though they can enable very fine-grained local control, such interaction modes are inefficient for the editing conditions that can be easily specified by language descriptions or reference images. Thanks to the recent breakthrough of cross-modal models (e.g., CLIP), HairCLIP is the first work that enables hair editing based on text descriptions or reference images. However, such text-driven and reference-driven interaction modes make HairCLIP unable to support fine-grained controls specified by sketch or mask. In this paper, we propose HairCLIPv2, aiming to support all the aforementioned interactions with one unified framework. Simultaneously, it improves upon HairCLIP with better irrelevant attributes (e.g., identity, background) preservation and unseen text descriptions support. The key idea is to convert all the hair editing tasks into hair transfer tasks, with editing conditions converted into different proxies accordingly. The editing effects are added upon the input image by blending the corresponding proxy features within the hairstyle or hair color feature spaces. Besides the unprecedented user interaction mode support, quantitative and qualitative experiments demonstrate the superiority of HairCLIPv2 in terms of editing effects, irrelevant attribute preservation and visual naturalness. Our code is available at https://github.com/wty-ustc/HairCLIPv2.

Recent advancements in 3D hand pose estimation have shown promising results, but its effectiveness has primarily relied on the availability of large-scale annotated datasets, the creation of which is a laborious and costly process. To alleviate the label-hungry limitation, we propose a self-supervised learning framework, HaMuCo, that learns a single view hand pose estimator from multi-view pseudo 2D labels. However, one of the main challenges of self-supervised learning is the presence of noisy labels and the "groupthink" effect from multiple views. To overcome these issues, we introduce a cross-view interaction network that distills the single view estimator by utilizing the cross-view correlated features and enforcing multi-view consistency to achieve collaborative learning. Both the single view estimator and the cross-view interaction network are trained jointly in an end-to-end manner. Extensive experiments show that our method can achieve state-of-the-art performance on multi-view self-supervised hand pose estimation. Furthermore, the proposed cross-view interaction network can also be applied to hand pose estimation from multi-view input and outperforms previous methods under same settings.

Referring Video Object Segmentation (RVOS) is to segment the object instance from a given video, according to the textual description of this object. However, in the open world, the object descriptions are often diversified in contents and flexible in lengths. This leads to the key difficulty in RVOS, i.e., various descriptions of different ob- jects are corresponding to different temporal scales in the video, which is ignored by most existing approaches with single stride of frame sampling. To tackle this problem, we propose a concise Hybrid Temporal-scale Multimodal Learning (HTML) framework, which can effectively align lingual and visual features to discover core object semantics in the video, by learning multimodal interaction hierarchically from different temporal scales. More specifically, we introduce a novel inter-scale multimodal perception module, where the language queries dynamically interact with visual features across temporal scales. It can effectively reduce complex object confusion by passing video context among different scales. Finally, we conduct extensive experiments on the widely used benchmarks, including Ref- Youtube-VOS, Ref-DAVIS17, A2D-Sentences and JHMDB- Sentences, where our HTML achieves state-of-the-art performance on all these datasets.

Despite the proven significance of hyperspectral images (HSIs) in performing various computer vision tasks, its potential is adversely affected by the low-resolution (LR) property in the spatial domain, resulting from multiple physical factors. Inspired by recent advancements in deep generative models, we propose an HSI Super-resolution (SR) approach with Conditional Diffusion Models (HSR-Diff) that merges a high-resolution (HR) multispectral image (MSI) with the corresponding LR-HSI. HSR-Diff generates an HR-HSI via repeated refinement, in which the HR-HSI is initialized with pure Gaussian noise and iteratively refined. At each iteration, the noise is removed with a Conditional Denoising Transformer (CDFormer) that is trained on denoising at different noise levels, conditioned on the hierarchical feature maps of HR-MSI and LR-HSI. In addition, a progressive learning strategy is employed to exploit the global information of full-resolution images. Systematic experiments have been conducted on four public datasets, demonstrating that HSR-Diff outperforms state-of-the-art methods.

Deep metric learning is crucial for finding an embedding function that can generalize to training and testing data, including unknown test classes. However, limited training samples restrict the model's generalization to downstream tasks. While adding new training samples is a promising solution, determining their labels remains a significant challenge. Here, we introduce Hybrid Species Embedding (HSE), which employs mixed sample data augmentations to generate hybrid species and provide additional training signals. We demonstrate that HSE outperforms multiple state-of-the-art methods in improving the metric Recall@K on the CUB-200 , CAR-196 and SOP datasets, thus offering a novel solution to deep metric learning's limitations.