ICCV 2023 Papers

Synthesizing high-fidelity videos from real-world multiview input is challenging due to the complexities of real-world environments and high-dynamic movements. Previous works based on neural radiance fields have demonstrated high-quality reconstructions of dynamic scenes. However, training such models on real-world scenes is time-consuming, usually taking days or weeks. In this paper, we present a novel method named MixVoxels to efficiently represent dynamic scenes which leads to fast training and rendering speed. The proposed MixVoxels represents the 4D dynamic scenes as a mixture of static and dynamic voxels and processes them with different networks. In this way, the computation of the required modalities for static voxels can be processed by a lightweight model, which essentially reduces the amount of computation as many daily dynamic scenes are dominated by static backgrounds. To distinguish the two kinds of voxels, we propose a novel variation field to estimate the temporal variance of each voxel. For the dynamic representations, we design an inner-product time query method to efficiently query multiple time steps, which is essential to recover the high-dynamic movements. As a result, with 15 minutes of training for dynamic scenes with inputs of 300-frame videos, MixVoxels achieves better PSNR than previous methods. For rendering, MixVoxels can render a novel view video with 1K resolution at 37 fps. Codes and trained models are available at https://github.com/fengres/mixvoxels.

Human pose estimation in videos has wide-ranging practical applications across various fields, many of which require fast inference on resource-scarce devices, necessitating the development of efficient and accurate algorithms. Previous works have demonstrated the feasibility of exploiting motion continuity to conduct pose estimation using sparsely sampled frames with transformer-based models. However, these methods only consider the temporal relation while neglecting spatial attention, and the complexity of dot product self-attention calculations in transformers are quadratically proportional to the embedding size. To address these limitations, we propose MixSynthFormer, a transformer encoder-like model with MLP-based mixed synthetic attention. By mixing synthesized spatial and temporal attentions, our model incorporates inter-joint and inter-frame importance and can accurately estimate human poses in an entire video sequence from sparsely sampled frames. Additionally, the flexible design of our model makes it versatile for other motion synthesis tasks. Our extensive experiments on 2D/3D pose estimation, body mesh recovery, and motion prediction validate the effectiveness and efficiency of MixSynthFormer.

Multi-media communications facilitate global interaction among people. However, despite researchers exploring cross-lingual translation techniques such as machine translation and audio speech translation to overcome language barriers, there is still a shortage of cross-lingual studies on visual speech. This lack of research is mainly due to the absence of datasets containing visual speech and translated text pairs. In this paper, we present AVMuST-TED, the first dataset for Audio-Visual Multilingual Speech Translation, derived from TED talks. Nonetheless, visual speech is not as distinguishable as audio speech, making it difficult to develop a mapping from source speech phonemes to the target language text. To address this issue, we propose MixSpeech, a cross-modality self-learning framework that utilizes audio speech to regularize the training of visual speech tasks. To further minimize the cross-modality gap and its impact on knowledge transfer, we suggest adopting mixed speech, which is created by interpolating audio and visual streams, along with a curriculum learning strategy to adjust the mixing ratio as needed. MixSpeech enhances speech translation in noisy environments, improving BLEU scores for four languages on AVMuST-TED by +1.4 to +4.2. Moreover, it achieves state-of-the-art performance in lip reading on CMLR (11.1%), LRS2 (25.5%), and LRS3 (28.0%).

Recently, semantic segmentation models trained with image-level text supervision have shown promising results in challenging open-world scenarios. However, these models still face difficulties in learning fine-grained semantic alignment at the pixel level and predicting accurate object masks. To address this issue, we propose MixReorg, a novel and straightforward pre-training paradigm for semantic segmentation that enhances a model's ability to reorganize patches mixed across images, exploring both local visual relevance and global semantic coherence. Our approach involves generating fine-grained patch-text pairs data by mixing image patches while preserving the correspondence between patches and text. The model is then trained to minimize the segmentation loss of the mixed images and the two contrastive losses of the original and restored features. With MixReorg as a mask learner, conventional text-supervised semantic segmentation models can achieve highly generalizable pixel-semantic alignment ability, which is crucial for open-world segmentation. After training with large-scale image-text data, MixReorg models can be applied directly to segment visual objects of arbitrary categories, without the need for further fine-tuning. Our proposed framework demonstrates strong performance on popular zero-shot semantic segmentation benchmarks, outperforming GroupViT by significant margins of 5.0%, 6.2%, 2.5%, and 3.4% mIoU on PASCAL VOC2012, PASCAL Context, MS COCO, and ADE20K, respectively.

Blending multiple convolutional kernels is proved advantageous in neural architecture design. However, current two-stage neural architecture search methods are mainly limited to single-path search spaces. How to efficiently search models of multi-path structures remains a difficult problem. In this paper, we are motivated to train a one-shot multi-path supernet to accurately evaluate the candidate architectures. Specifically, we discover that in the studied search spaces, feature vectors summed from multiple paths are nearly multiples of those from a single path. Such disparity perturbs the supernet training and its ranking ability. Therefore, we propose a novel mechanism called Shadow Batch Normalization (SBN) to regularize the disparate feature statistics. Extensive experiments prove that SBNs are capable of stabilizing the optimization and improving ranking performance. We call our unified multi-path one-shot approach as MixPath, which generates a series of models that achieve state-of-the-art results on ImageNet.

3D single object tracking (SOT) is an indispensable part of automated driving. Existing approaches rely heavily on large, densely labeled datasets. However, annotating point clouds is both costly and time-consuming. Inspired by the great success of cycle tracking in unsupervised 2D SOT, we introduce the first semi-supervised approach to 3D SOT. Specifically, we introduce two cycle-consistency strategies for supervision: 1) Self tracking cycles, which leverage labels to help the model converge better in the early stages of training; 2) forward-backward cycles, which strengthen the tracker's robustness to motion variations and the template noise caused by the template update strategy. Furthermore, we propose a data augmentation strategy named SOTMixup to improve the tracker's robustness to point cloud diversity. SOTMixup generates training samples by sampling points in two point clouds with a mixing rate and assigns a reasonable loss weight for training according to the mixing rate. The resulting MixCycle approach generalizes to appearance matching-based trackers. On the KITTI benchmark, based on the P2B tracker, MixCycle trained with 10% labels outperforms P2B trained with 100% labels, and achieves a 28.4% precision improvement when using 1% labels. Our code will be released at https://github.com/Mumuqiao/MixCycle.

Learning from label proportions (LLP) is a promising weakly supervised learning problem. In LLP, a set of instances (bag) has label proportions but no instance-level labels. LLP aims to train an instance-level classifier by using the label proportions of the bag. In this paper, we propose a bag-level data augmentation method for LLP called MixBag, which is based on the key observation from our preliminary experiments; that the instance-level classification accuracy improves as the number of labeled bags increases even though the total number of instances is fixed. We also propose a confidence interval loss designed based on statistical theory in order to use the augmented bags effectively. To the best of our knowledge, this is the first attempt to propose bag-level data augmentation for LLP. The advantage of MixBag is that it can be applied to instance-level data augmentation techniques and any LLP method that uses the proportion loss. Experimental results demonstrate this advantage and the effectiveness of our method.

In video action recognition, shortcut static features can interfere with the learning of motion features, resulting in poor out-of-distribution (OOD) generalization. The video background is clearly a source of static bias, but the video foreground, such as the clothing of the actor, can also provide static bias. In this paper, we empirically verify the existence of foreground static bias by creating test videos with conflicting signals from the static and moving portions of the video. To tackle this issue, we propose a simple yet effective technique, StillMix, to learn robust action representations. Specifically, StillMix identifies bias-inducing video frames using a 2D reference network and mixes them with videos for training, serving as effective bias suppression even when we cannot explicitly extract the source of bias within each video frame or enumerate types of bias. Finally, to precisely evaluate static bias, we synthesize two new benchmarks, SCUBA for static cues in the background, and SCUFO for static cues in the foreground. With extensive experiments, we demonstrate that StillMix mitigates both types of static bias and improves video representations for downstream applications.

Adversarial examples derived from deliberately crafted perturbations on visual inputs can easily harm decision process of deep neural networks. To prevent potential threats, various adversarial training-based defense methods have grown rapidly and become a de facto standard approach for robustness. Despite recent competitive achievements, we observe that adversarial vulnerability varies across targets and certain vulnerabilities remain prevalent. Intriguingly, such peculiar phenomenon cannot be relieved even with deeper architectures and advanced defense methods. To address this issue, in this paper, we introduce a causal approach called Adversarial Double Machine Learning (ADML), which allows us to quantify the degree of adversarial vulnerability for network predictions and capture the effect of treatments on outcome of interests. ADML can directly estimate causal parameter of adversarial perturbations per se and mitigate negative effects that can potentially damage robustness, bridging a causal perspective into the adversarial vulnerability. Through extensive experiments on various CNN and Transformer architectures, we corroborate that ADML improves adversarial robustness with large margins and relieve the empirical observation.

Contrastive Language-Image Pretraining has emerged as a prominent approach for training vision and text encoders with uncurated image-text pairs from the web. To enhance data-efficiency, recent efforts have introduced additional supervision terms that involve random-augmented views of the image. However, since the image augmentation process is unaware of its text counterpart, this procedure could cause various degrees of image-text misalignments during training. Prior methods either disregarded this discrepancy or introduced external models to mitigate the impact of misalignments during training. In contrast, we propose a novel metric learning approach that capitalizes on these misalignments as an additional training source, which we term "Misalign, Contrast then Distill (MCD)". Unlike previous methods that treat augmented images and their text counterparts as simple positive pairs, MCD predicts the continuous scales of misalignment caused by the augmentation. Our extensive experimental results show that our proposed MCD achieves state-of-the-art transferability in multiple classification and retrieval downstream datasets.

Despite significant research efforts, deep neural networks remain vulnerable to biases: this raises concerns about their fairness and limits their generalization. In this paper, we propose a bias-agnostic approach to mitigate the impact of biases in deep neural networks. Unlike traditional debiasing approaches, we rely on a metric to quantify "bias alignment/misalignment" on target classes and use this information to discourage the propagation of bias-target alignment information through the network. We conduct experiments on several commonly used datasets for debiasing and compare our method with supervised and bias-specific approaches. Our results indicate that the proposed method achieves comparable performance to state-of-the-art supervised approaches, despite being bias-agnostic, even in the presence of multiple biases in the same sample.

We present a novel camera path optimization framework for the task of online video stabilization. Typically, a stabilization pipeline consists of three steps: motion estimating, path smoothing, and novel view rendering. Most previous methods concentrate on motion estimation, proposing various global or local motion models. In contrast, path optimization receives relatively less attention, especially in the important online setting, where no future frames are available. In this work, we adopt recent off-the-shelf high-quality deep motion models for motion estimation to recover the camera trajectory and focus on the latter two steps. Our network takes a short 2D camera path in a sliding window as input and outputs the stabilizing warp field of the last frame in the window, which warps the coming frame to its stabilized position. A hybrid loss is well-defined to constrain the spatial and temporal consistency. In addition, we build a motion dataset that contains stable and unstable motion pairs for the training. Extensive experiments demonstrate that our approach significantly outperforms state-of-the-art online methods both qualitatively and quantitatively and achieves comparable performance to offline methods.

This paper proposes minimal solutions to uncalibrated two-view geometry with known epipoles. Exploiting the epipoles, we can reduce the number of point correspondences needed to find the fundamental matrix together with the intrinsic parameters: the focal length and the radial lens distortion. We define four cases by the number of available epipoles and unknown intrinsic parameters, then derive a closed-form solution for each case formulated as a higher-order polynomial in a single variable. The proposed solvers are more numerically stable and faster by orders of magnitude than the conventional 6- or 7-point algorithms. Moreover, we demonstrate by experiments on the human pose dataset that the proposed method can solve two-view geometry even with 2D human pose, of which point localization is noisier than general feature point detectors.

For a generalized (or non-central) camera model, the minimal problem for two views of six points has efficient solvers. However, minimal problems of three views with four points and three views of six lines have not yet been explored and solved, despite the efforts from the computer vision community. This paper develops the formulations of these two minimal problems and shows how state-of-the-art GPU implementations of Homotopy Continuation solver can be used effectively. The proposed methods are evaluated on both synthetic and real datasets, demonstrating that they are fast, accurate and that they improve on structure from motion estimations, when employed in an hypothesis and test setting.

Online Action Detection (OAD) is the task of identifying actions in streaming videos without access to future frames. Much effort has been devoted to effectively capturing long-range dependencies, with transformers receiving the spotlight for their ability to capture long-range temporal structures. In contrast, RNNs have received less attention lately, due to their lower performance compared to recent methods that utilize transformers. In this paper, we investigate the underlying reasons for the inferior performance of RNNs compared to transformer-based algorithms. Our findings indicate that the discrepancy between training and inference is the primary hindrance to the effective training of RNNs. To address this, we propose applying non-uniform weights to the loss computed at each time step, which allows the RNN model to learn from the predictions made in an environment that better resembles the inference stage. Extensive experiments on three benchmark datasets, THUMOS, TVSeries, and FineAction demonstrate that a minimal RNN-based model trained with the proposed methodology performs equally or better than the existing best methods with a significant increase in efficiency. The code is available at https://github.com/jbistanbul/MiniROAD.

Generating images with both photorealism and multiview 3D consistency is crucial for 3D-aware GANs, yet existing methods struggle to achieve them simultaneously. Improving the photorealism via CNN-based 2D super-resolution can break the strict 3D consistency, while keeping the 3D consistency by learning high-resolution 3D representations for direct rendering often compromises image quality. In this paper, we propose a novel learning strategy, namely 3D-to-2D imitation, which enables a 3D-aware GAN to generate high-quality images while maintaining their strict 3D consistency, by letting the images synthesized by the generator's 3D rendering branch mimic those generated by its 2D super-resolution branch. We also introduce 3D-aware convolutions into the generator for better 3D representation learning, which further improves the image generation quality. With the above strategies, our method reaches FID scores of 5.4 and 4.3 on FFHQ and AFHQ-v2 Cats, respectively, at 512x512 resolution, largely outperforming existing 3D-aware GANs using direct 3D rendering and coming very close to the previous state-of-the-art method that leverages 2D super-resolution. Project website: https://seanchenxy.github.io/Mimic3DWeb.

Reconstructing accurate 3D scenes from images is a long-standing vision task. Due to the ill-posedness of the single-image reconstruction problem, most well-established methods are built upon multi-view geometry. State-of-the-art (SOTA) monocular metric depth estimation methods can only handle a single camera model and are unable to perform mixed-data training due to the metric ambiguity. Meanwhile, SOTA monocular methods trained on large mixed datasets achieve zero-shot generalization by learning affine-invariant depths, which cannot recover real-world metrics. In this work, we show that the key to a zero-shot single-view metric depth model lies in the combination of large-scale data training and resolving the metric ambiguity from various camera models. We propose a canonical camera space transformation module, which explicitly addresses the ambiguity problems and can be effortlessly plugged into existing monocular models. Equipped with our module, monocualr models can be stably trained over 8 millions of images with thousands of camera models, resulting in zero-shot generalization to in-the-wild images with unseen camera settings. Experiments demonstrate SOTA performance of our method on 7 zero-shot benchmarks. Our method can recover the metric 3D structure on randomly collected Internet images, enabling plausible single-image metrology. Downstream tasks can also be significantly improved by naively plug-in our model. E.g., our model relieves the scale drift issues of monocular-SLAM (Fig. 1), leading to metric scale high-quality dense mapping.

In this paper, we consider a real-world scenario where a model that is trained on pre-defined classes continually encounters unlabeled data that contains both known and novel classes. The goal is to continually discover novel classes while maintaining the performance in known classes. We name the setting Continual Generalized Category Discovery (C-GCD). Existing methods for novel class discovery cannot directly handle the C-GCD setting due to some unrealistic assumptions, such as the unlabeled data only containing novel classes. Furthermore, they fail to discover novel classes in a continual fashion. In this work, we lift all these assumptions and propose an approach, called MetaGCD, to learn how to incrementally discover with less forgetting. Our proposed method uses a meta-learning framework and leverages the offline labeled data to simulate the testing incremental learning process. A meta-objective is defined to revolve around two conflicting learning objectives to achieve novel class discovery without forgetting. Furthermore, a soft neighborhood-based contrastive network is proposed to discriminate uncorrelated images while attracting correlated images. We build strong baselines and conduct extensive experiments on three widely used benchmarks to demonstrate the superiority of our method.

Due to their highly structured characteristics, faces are easier to recover than natural scenes for blind image super-resolution. Therefore, we can extract the degradation representation of an image from the low-quality and recovered face pairs. Using the degradation representation, realistic low-quality images can then be synthesized to fine-tune the super-resolution model for the real-world low-quality image. However, such a procedure is time-consuming and laborious, and the gaps between recovered faces and the ground-truths further increase the optimization uncertainty. To facilitate efficient model adaptation towards image-specific degradations, we propose a method dubbed MetaF2N, which leverages the contained faces to fine-tune model parameters for adapting to the whole natural image in a meta-learning framework. The degradation extraction and low-quality image synthesis steps are thus circumvented in our MetaF2N, and it requires only one fine-tuning step to get decent performance. Considering the gaps between the recovered faces and ground-truths, we further deploy a MaskNet for adaptively predicting loss weights at different positions to reduce the impact of low-confidence areas. To evaluate our proposed MetaF2N, we have collected a real-world low-quality dataset with one or multiple faces in each image, and our MetaF2N achieves superior performance on both synthetic and realworld datasets. Source code, pre-trained models, and collected datasets are available at https://github.com/yinzhicun/MetaF2N.