ECCV 2022 Papers

"Current supervised cross-domain image retrieval methods can achieve excellent performance. However, the cost of data collection and labeling imposes an intractable barrier to practical deployment in real applications. In this paper, we investigate the unsupervised cross-domain image retrieval task, where class labels and pairing annotations are no longer a prerequisite for training. This is an extremely challenging task because there is no supervision for both in-domain feature representation learning and cross-domain alignment. We address both challenges by introducing: 1) a new cluster-wise contrastive learning mechanism to help extract class semantic-aware features, and 2) a novel distance-of-distance loss to effectively measure and minimize the domain discrepancy without any external supervision. Experiments on the Office-Home and DomainNet datasets consistently show the superior image retrieval accuracies of our framework over state-of-the-art approaches. Our source code can be found at https://github.com/conghuihu/UCDIR."

"While the voxel-based methods have achieved promising results for multi-person 3D pose estimation from multi-cameras, they suffer from heavy computation burdens, especially for large scenes. We present Faster VoxelPose to address the challenge by re-projecting the feature volume to the three two-dimensional coordinate planes and estimating X, Y, Z coordinates from them separately. To that end, we first localize each person by a 3D bounding box by estimating a 2D box and its height based on the volume features projected to the xy-plane and z-axis, respectively. Then for each person, we estimate partial joint coordinates from the three coordinate planes separately which are then fused to obtain the final 3D pose. The method is free from costly 3D-CNNs and improves the speed of VoxelPose by ten times and meanwhile achieves competitive accuracy as the state-of-the-art methods, proving its potential in real-time applications."

"Video-to-Video synthesis (Vid2Vid) has achieved remarkable results on generating a photo-realistic video from a sequence of semantic maps. However, this pipeline suffers from high computational cost and long inference latency, which largely depends on two essential factors: 1) network architecture parameters, 2) sequential data stream. Recently, the parameters of image-based generative models have been significantly compressed via more efficient network architectures. Nevertheless, existing methods mainly focus on slimming network architectures and ignore the size of sequential data stream. Moreover, due to the lack of temporal coherence, image-based compression is not sufficient for the compression of video synthesis with a dynamic time series. In this paper, we present a spatial-temporal compression framework, Fast-Vid2Vid, which focuses on data aspects of generative models. It makes the first attempt at time dimension to reduce computational resources and accelerate inference. Specifically, we compress the input data stream spatially and reduce the temporal redundancy. After the proposed spatial-knowledge distillation, our model can synthesize key-frames using low-resolution data stream. Finally, Fast-Vid2Vid interpolates inter frames by motion compensation with negligible latency. On standard benchmarks, Fast-Vid2Vid achieves around real-time performance as 20 FPS and saves around 8× computational cost on a single V100 GPU."

"Contrastive-based self-supervised learning methods achieved great success in recent years. However, self-supervision requires extremely long training epochs (e.g., 800 epochs for MoCo v3) to achieve promising results, which is unacceptable for the general academic community and hinders the development of this topic. This work revisits the momentum-based contrastive learning frameworks and identifies the inefficiency in which two augmented views generate only one positive pair. We propose Fast-MoCo - a novel framework that utilizes combinatorial patches to construct multiple positive pairs from two augmented views, which provides abundant supervision signals that bring significant acceleration with neglectable extra computational cost. Fast-MoCo trained with 100 epochs achieves 73.5% linear evaluation accuracy, similar to MoCo v3 (ResNet-50 backbone) trained with 800 epochs. Extra training (200 epochs) further improves the result to 75.1%, which is on par with state-of-the-art methods. Experiments on several downstream tasks also confirm the effectiveness of Fast-MoCo."

"Most image denoising networks apply a single set of static convolutional kernels across the entire input image. This is sub-optimal for natural images, as they often consist of heterogeneous visual patterns. Dynamic convolution tries to address this issue by using per-pixel convolution kernels, but this greatly increases computational cost. In this work, we present \textbf{Malle}able \textbf{Conv}olution (\textbf{MalleConv}), which performs spatial-varying processing with minimal computational overhead. MalleConv uses a smaller set of spatially-varying convolution kernels, a compromise between static and per-pixel convolution kernels. These spatially-varying kernels are produced by an efficient predictor network running on a downsampled input, making them much more efficient to compute than per-pixel kernels produced by a full-resolution image, and also enlarging the network’s receptive field compared with static kernels. These kernels are then jointly upsampled and applied to a full-resolution feature map through an efficient on-the-fly slicing operator with minimum memory overhead. To demonstrate the effectiveness of MalleConv, we use it to build an efficient denoising network we call \textbf{MalleNet}. MalleNet achieves high quality results without a very deep architecture, e.g., running 8.9$\times$ faster than the best performing denoising algorithms (SwinIR) while maintaining similar quality. We also show that a single MalleConv layer added to a standard convolution-based backbone can contribute significantly to reducing the computational cost or can boost image quality at a similar cost."

"We address the problem of segmenting moving rigid objects based on two-view image correspondences under a perspective camera model. While this is a well understood problem, existing methods scale poorly with the number of correspondences. In this paper we propose a fast segmentation algorithm that scales linearly with the number of correspondences and show that on benchmark datasets it offers the best trade-off between error and computational time: it is at least one order of magnitude faster than the best method (with comparable or better accuracy), with the ratio growing up to three orders of magnitude for larger number of correspondences. We approach the problem from an algebraic perspective by exploiting the fact that all points belonging to a given object lie in the same quadratic surface. The proposed method is based on a characterization of each surface in terms of the Christoffel polynomial associated with the probability that a given point belongs to the surface. This allows for efficiently segmenting points “one surface at a time” in O(number of points)."

"The optics of any camera degrades the sharpness of photographs, which is a key visual quality criterion. This degradation is characterized by the point-spread function (PSF), which depends on the wavelengths of light and is variable across the imaging field. In this paper, we propose a two-step scheme to correct optical aberrations in a single raw or JPEG image, i.e., without any prior information on the camera or lens. First, we estimate local Gaussian blur kernels for overlapping patches and sharpen them with a non-blind deblurring technique. Based on the measurements of the PSFs of dozens of lenses, these blur kernels are modeled as RGB Gaussians defined by seven parameters. Second, we remove the remaining lateral chromatic aberrations (not contemplated in the first step) with a convolutional neural network, trained to minimize the red/green and blue/green residual images. Experiments on both synthetic and real images show that the combination of these two stages yields a fast state-of-the-art blind optical aberration compensation technique that competes with commercial non-blind algorithms."

"Large-scale Vision-and-Language (V+L) pre-training for representation learning has proven to be effective in boosting various downstream V+L tasks. However, when it comes to the fashion domain, existing V+L methods are inadequate as they overlook the unique characteristics of both fashion V+L data and downstream tasks. In this work, we propose a novel fashion-focused V+L representation learning framework, dubbed as FashionViL. It contains two novel fashion-specific pre-training tasks designed particularly to exploit two intrinsic attributes with fashion V+L data. First, in contrast to other domains where a V+L datum contains only a single image-text pair, there could be multiple images in the fashion domain. We thus propose a Multi-View Contrastive Learning task for pulling closer the visual representation of one image to the compositional multimodal representation of another image+text. Second, fashion text (e.g., product description) often contains rich fine-grained concepts (attributes/noun phrases). To capitalize this, a Pseudo-Attributes Classification task is introduced to encourage the learned unimodal (visual/textual) representations of the same concept to be adjacent. Further, fashion V+L tasks uniquely include ones that do not conform to the common one-stream or two-stream architectures (e.g., text-guided image retrieval). We thus propose a flexible, versatile V+L model architecture consisting of a modality-agnostic Transformer so that it can be flexibly adapted to any downstream tasks. Extensive experiments show that our FashionViL achieves new state of the art across five downstream tasks. Code is available at https://github.com/BrandonHanx/mmf."

"Data-Efficient GANs (DE-GANs), which aim to learn generative models with a limited amount of training data, encounter several challenges for generating high-quality samples. Since data augmentation strategies have largely alleviated the training instability, how to further improve the generative performance of DE-GANs becomes a hotspot. Recently, contrastive learning has shown the great potential of increasing the synthesis quality of DE-GANs, yet related principles are not well explored. In this paper, we revisit and compare different contrastive learning strategies in DE-GANs, and identify (i) the current bottleneck of generative performance is the discontinuity of latent space; (ii) compared to other contrastive learning strategies, Instance-perturbation works towards latent space continuity, which brings the major improvement to DE-GANs. Based on these observations, we propose FakeCLR, which only applies contrastive learning on perturbed fake samples, and devises three related training techniques: Noise-related Latent Augmentation, Diversity-aware Queue, and Forgetting Factor of Queue. Our experimental results manifest the new state of the arts on both few-shot generation and limited-data generation. On multiple datasets, FakeCLR acquires more than 15% FID improvement compared to existing DE-GANs. Code is available at https://github.com/iceli1007/FakeCLR."

"Recent advances in generative adversarial networks have shown that it is possible to generate high-resolution and hyperrealistic images. However, the images produced by GANs are only as fair and representative as the datasets on which they are trained. In this paper, we propose a method for directly modifying a pre-trained StyleGAN2 model that can be used to generate a balanced set of images with respect to one (e.g., eyeglasses) or more attributes (e.g., gender and eyeglasses). Our method takes advantage of the style space of the StyleGAN2 model to perform disentangled control of the target attributes to be debiased. Our method does not require training additional models and directly debiases the GAN model, paving the way for its use in various downstream applications. Our experiments show that our method successfully debiases the GAN model within a few minutes without compromising the quality of the generated images. To promote fair generative models, we share the code and debiased models at http://catlab-team.github.io/fairstyle."

"Existing pruning techniques preserve deep neural networks’ overall ability to make correct predictions but could also amplify hidden biases during the compression process. We propose a novel pruning method, Fairness-aware GRAdient Pruning mEthod (FairGRAPE), that minimizes the disproportionate impacts of pruning on different sub-groups. Our method calculates the per-group importance of each model weight and selects a subset of weights that maintain the relative between-group total importance in pruning. The proposed method then prunes network edges with small importance values and repeats the procedure by updating importance values. We demonstrate the effectiveness of our method on four different datasets, FairFace, UTKFace, CelebA, and ImageNet, for the tasks of face attribute classification where our method reduces the disparity in performance degradation by up to 90% compared to the state-of-the-art pruning algorithms. Our method is substantially more effective in a setting with a high pruning rate (99%). The code and dataset used in the experiments are available at https://github.com/Bernardo1998/FairGRAPE"

"In this paper, we explore a novel and ambitious knowledge-transfer task, termed Knowledge Factorization (KF). The core idea of KF lies in the modularization and assemblability of knowledge: given a pretrained network model as input, KF aims to decompose it into several factor networks, each of which handles only a dedicated task and maintains task-specific knowledge factorized from the source network. Such factor networks are task-wise disentangled and can be directly assembled, without any fine-tuning, to produce the more competent combined-task networks. In other words, the factor networks serve as Lego-brick-like building blocks, allowing us to construct customized networks in a plug-and-play manner. Specifically, each factor network comprises two modules, a common-knowledge module that is task-agnostic and shared by all factor networks, alongside with a task-specific module dedicated to the factor network itself. We introduce an information-theoretic objective, InfoMax-Bottleneck (IMB), to carry out KF by optimizing the mutual information between the learned representations and input. Experiments across various benchmarks demonstrate that, the derived factor networks yield gratifying performances on not only the dedicated tasks but also disentanglement, while enjoying much better interpretability and modularity. Moreover, the learned common-knowledge representations give rise to impressive results on transfer learning. Our code is available at https://github.com/Adamdad/KnowledgeFactor."

"Recently, Dual-Pixel (DP) sensors have been adopted in many imaging devices. However, despite their various advantages, DP sensors are used just for faster auto-focus and aesthetic image captures, and research on their usage for 3D facial understanding has been limited due to the lack of datasets and algorithmic designs that exploit parallax in DP images. It is also because the baseline of sub-aperture images is extremely narrow, and parallax exists in the defocus blur region. In this paper, we introduce a DP-oriented Depth/Normal estimation network that reconstructs the 3D facial geometry. In addition, to train the network, we collect DP facial data with more than 135K images for 101 persons captured with our multi-camera structured light systems. It contains ground-truth 3D facial models including depth map and surface normal in metric scale. Our dataset allows the proposed network to be generalized for 3D facial depth/normal estimation. The proposed network consists of two novel modules: Adaptive Sampling Module (ASM) and Adaptive Normal Module (ANM), which are specialized in handling the defocus blur in DP images. Finally, we demonstrate that the proposed method achieves state-of-the-art performances over recent DP-based depth/normal estimation methods."

"Existing one-shot face reenactment methods either present obvious artifacts in large pose transformations, or cannot well-preserve the identity information in the source images, or fail to meet the requirements of real-time applications due to the intensive amount of computation involved. In this paper, we introduce Face2Face^ρ, the first Real-time High-resolution and One-shot (RHO, ρ) face reenactment framework. To achieve this goal, we designed a new 3DMM-assisted warping-based face reenactment architecture which consists of two fast and efficient sub-networks, i.e., a u-shaped rendering network to reenact faces driven by head poses and facial motion fields, and a hierarchical coarse-to-fine motion network to predict facial motion fields guided by different scales of landmark images. Compared with existing state-of-the-art works, Face2Face^ρ can produce results of equal or better visual quality, yet with significantly less time and memory overhead. We also demonstrate that Face2Face^ρ can achieve real-time performance for face images of 1440×1440 resolution with a desktop GPU and 256×256 resolution with a mobile CPU."

"Fabric materials are central to recreating realistic appearance of avatars in a virtual world and many VR applications, ranging from virtual try-on, teleconferencing, to character animation. We propose an end-to-end network model that uses video input to estimate the fabric materials of the garment worn by a human or an avatar in a virtual world. To achieve the high accuracy, we jointly learn human body and the garment geometry as conditions to material prediction. Due to the highly dynamic and deformable nature of cloth, general data-driven garment modeling remains a challenge. To address this problem, we propose a two-level auto-encoder to account for both global and local features of any garment geometry that would directly affect material perception. Using this network, we can also achieve smooth geometry transitioning between different garment topologies. During the estimation, we use a closed-loop optimization structure to share information between tasks and feed the learned garment features for temporal estimation of garment materials. Experiments show that our proposed network structures greatly improve the material classification accuracy by 1.5x, with applicability to unseen input. It also runs at least three orders of magnitude faster than the state-of-the-art. We demonstrate the recovered fabric materials on virtual try-on, where we recreate the entire avatar appearance, including body shape and pose, garment geometry and materials from only a single video."

"We advance sketch research to scenes with the first dataset of freehand scene sketches, FSCOCO. With practical applications in mind, we collect sketches that convey well scene content but can be sketched within a few minutes by a person with any sketching skills. Our dataset comprises 10,000 freehand scene vector sketches with per point space-time information by 100 non-expert individuals, offering both object- and scene-level abstraction. Each sketch is augmented with its text description. Using our dataset, we study for the first time the problem of fine-grained image retrieval from freehand scene sketches and sketch captions. We draw insights on: (i) Scene salience encoded in sketches using the strokes temporal order; (ii) Performance comparison of image retrieval from a scene sketch and an image caption; (iii) Complementarity of information in sketches and image captions, as well as the potential benefit of combining the two modalities. In addition, we extend a popular vector sketch LSTM-based encoder to handle sketches with larger complexity than was supported by previous work. Namely, we propose a hierarchical sketch decoder, which we leverage at a sketch-specific “pretext” task. Our dataset enables for the first time research on freehand scene sketch understanding and its practical applications. We will release the dataset upon acceptance."

"The ability to learn new concepts continually is necessary in this ever-changing world. However, deep neural networks suffer from catastrophic forgetting when learning new categories. Many works have been proposed to alleviate this phenomenon, whereas most of them either fall into the stability-plasticity dilemma or take too much computation or storage overhead. Inspired by the gradient boosting algorithm to gradually fit the residuals between the target model and the previous ensemble model, we propose a novel two-stage learning paradigm FOSTER, empowering the model to learn new categories adaptively. Specifically, we first dynamically expand new modules to fit the residuals between the target and the output of the original model. Next, we remove redundant parameters and feature dimensions through an effective distillation strategy to maintain the single backbone model. We validate our method FOSTER on CIFAR-100 and ImageNet-100/1000 under different settings. Experimental results show that our method achieves state-of-the-art performance. Code is available at https://github.com/G-U-N/ECCV22-FOSTER."

"The increasing availability of video recordings made by multiple cameras has offered new means for mitigating occlusion and depth ambiguities in pose and motion reconstruction methods. Yet, multi-view algorithms strongly depend on camera parameters, particularly on relative transformations between the cameras. Such a dependency becomes a hurdle once shifting to dynamic capture in uncontrolled settings. We introduce FLEX (Free muLti-view rEconstruXion), an end-to-end extrinsic parameter-free multi-view model. FLEX is extrinsic parameter-free (dubbed ep-free) in the sense that it does not require extrinsic camera parameters. Our key idea is that the 3D angles between skeletal parts, as well as bone lengths, are invariant to the camera position. Hence, learning 3D rotations and bone lengths rather than locations allows for predicting common values for all camera views. Our network takes multiple video streams, learns fused deep features through a novel multi-view fusion layer, and reconstructs a single consistent skeleton with temporally coherent joint rotations. We demonstrate quantitative and qualitative results on three public data sets, and on multi-person synthetic video streams captured by dynamic cameras. We compare our model to state-of-the-art methods that are not ep-free and show that in the absence of camera parameters, we outperform them by a large margin while obtaining comparable results when camera parameters are available. Code, trained models, and other materials are available on https://briang13.github.io/FLEX."

"We present a frame interpolation algorithm that synthesizes an engaging slow-motion video from near-duplicate photos which often exhibit large scene motion. Near-duplicates interpolation is an interesting new application, but large motion poses challenges to existing methods. To address this issue, we adapt a feature extractor that shares weights across the scales, and present a ""scale-agnostic"" motion estimator. It relies on the intuition that large motion at finer scales should be similar to small motion at coarser scales, which boosts the number of available pixels for large motion supervision. To inpaint wide disocclusions caused by large motion and synthesize crisp frames, we propose to optimize our network with the Gram matrix loss that measures the correlation difference between features. To simplify the training process, we further propose a unified single-network approach that removes the reliance on additional optical-flow or depth network and is trainable from frame triplets alone. Our approach outperforms state-of-the-art methods on the Xiph large motion benchmark while performing favorably on Vimeo-90K, Middlebury and UCF101. Codes and pre-trained models are available at https://film-net.github.io."