ECCV 2022 Papers

"Successful depth completion from a single RGB-D image requires both extracting plentiful 2D and 3D features and merging these heterogeneous features appropriately. We propose a novel depth completion framework, CostDCNet, based on the cost volume-based depth estimation approach that has been successfully employed for multi-view stereo (MVS). The key to high-quality depth map estimation in the approach is constructing an accurate cost volume. To produce a quality cost volume tailored to single-view depth completion, we present a simple but effective architecture that can fully exploit the 3D information, three options to make an RGB-D feature volume, and per-plane pixel shuffle for efficient volume upsampling. Our CostDCNet framework consists of lightweight deep neural networks ( 1.8M parameters), running in real time ( 30ms). Nevertheless, thanks to our simple but effective design, CostDCNet demonstrates depth completion results comparable to or better than the state-of-the-art methods."

"We present a novel cost aggregation network, called Volumetric Aggregation with Transformers (VAT), for few-shot segmentation. The use of transformers can benefit correlation map aggregation through self-attention over a global receptive field. However, the tokenization of a correlation map for transformer processing can be detrimental, because the discontinuity at token boundaries reduces the local context available near the token edges and decreases inductive bias. To address this problem, we propose a 4D Convolutional Swin Transformer, where a high-dimensional Swin Transformer is preceded by a series of small-kernel convolutions that impart local context to all pixels and introduce convolutional inductive bias. We additionally boost aggregation performance by applying transformers within a pyramidal structure, where aggregation at a coarser level guides aggregation at a finer level. Error in the transformer output is then filtered in the subsequent decoder with the help of the query’s appearance embedding. With this model, a new state-of-the-art is set for all the standard benchmarks in few-shot segmentation. It is shown that VAT attains state-of-the-art performance for semantic correspondence as well, where cost aggregation also plays a central role."

"Translation averaging methods use the consistency of input translation directions to solve for camera translations. However, translation directions obtained using epipolar geometry are error-prone. This paper argues that the improved accuracy of translation averaging should be leveraged to mitigate the errors in the input translation direction estimates. To this end, we introduce weights for individual correspondences which are iteratively refined to yield improved translation directions. In turn, these refined translation directions are averaged to obtain camera translations. This results in an alternating approach to translation averaging. The modularity of our framework allows us to use existing translation averaging methods and improve their results. The efficacy of the scheme is demonstrated by comparing performance with state-of-the-art methods on a number of real-world datasets. We also show that our approach yields reasonably good 3D reconstructions with straightforward triangulation, i.e. without any bundle adjustment iterations."

"Corner-based object detectors enjoy the potential of detecting arbitrarily-sized instances, yet the performance is mainly harmed by the accuracy of instance construction. Specifically, there are three factors, namely, 1) the corner keypoints are prone to false-positives; 2) incorrect matches emerge upon corner keypoint pull-push embeddings; and 3) the heuristic NMS cannot adjust the corners pull-push mechanism. Accordingly, this paper presents an elegant framework named Cornerformer that is composed of two factors. First, we build a Corner Transformer Encoder (CTE, a self-attention module) in a 2D-form to enhance the information aggregated by corner keypoints, offering stronger features for the pull-push loss to distinguish instances from each other. Second, we design an Attenuation-Auto-Adjusted NMS (A3-NMS) to maximally leverage the semantic outputs and avoid true objects from being removed. Experiments on object detection and human pose estimation show the superior performance of Cornerformer in terms of accuracy and inference speed."

"Vision Transformers (ViTs) have recently dominated a range of computer vision tasks, yet it suffers from low training data efficiency and inferior local semantic representation capability without appropriate inductive bias. Convolutional neural networks (CNNs) inherently capture regional-aware semantics, inspiring researchers to introduce CNNs back into the architecture of the ViTs to provide desirable inductive bias for ViTs. However, is the locality achieved by the micro-level CNNs embedded in ViTs good enough? In this paper, we investigate the problem by profoundly exploring how the macro architecture of the hybrid CNNs/ViTs enhances the performances of hierarchical ViTs. Particularly, we study the role of token embedding layers, alias convolutional embedding (CE), and systemically reveal how CE injects desirable inductive bias in ViTs. Besides, we apply the optimal CE configuration to 4 recently released state-of-the-art ViTs, effectively boosting the corresponding performances. Finally, a family of efficient hybrid CNNs/ViTs, dubbed CETNets, are released, which may serve as generic vision backbones. Specifically, CETNets achieve 84.9% Top-1 accuracy on ImageNet-1K (training from scratch), 48.6% box mAP on the COCO benchmark, and 51.6% mIoU on the ADE20K, substantially improving the performances of the corresponding state-of-the-art baselines."

"Although significant advances have been made in face recognition (FR), FR in unconstrained environments remains challenging due to the domain gap between the semi-constrained training datasets and unconstrained testing scenarios. To address this problem, we propose a controllable face synthesis model (CFSM) that can mimic the distribution of target datasets in a style latent space. CFSM learns a linear subspace with orthogonal bases in the style latent space with precise control over the diversity and degree of synthesis. Furthermore, the pre-trained synthesis model can be guided by the FR model, making the resulting images more beneficial for FR model training. Besides, target dataset distributions are characterized by the learned orthogonal bases, which can be utilized to measure the distributional similarity among face datasets. Our approach yields significant performance gains on unconstrained benchmarks, such as IJB-B, IJB-C, TinyFace and IJB-S (+5.76% Rank1)."

"We present GLASS, a method for Global and Local Action-driven Sequence Synthesis. GLASS is a generative model that is trained on video sequences in an unsupervised manner and that can animate an input image at test time. The method learns to segment frames into foreground-background layers and to generate transitions of the foregrounds over time through a global and local action representation. Global actions are explicitly related to 2D shifts, while local actions are instead related to (both geometric and photometric) local deformations. GLASS uses a recurrent neural network to transition between frames and is trained through a reconstruction loss. We also introduce W-Sprites (Walking Sprites), a novel synthetic dataset with a predefined action space. We evaluate our method on both W-Sprites and real datasets, and find that GLASS is able to generate realistic video sequences from a single input image and to successfully learn a more advanced action space than in prior work. Further details, the code and example videos are available at https://araachie.github.io/glass/."

"Shadows are essential for realistic image compositing. Physics based shadow rendering methods require 3D geometries, which are not always available. Deep learning-based shadow synthesis methods learn a mapping from the light information to an object’s shadow without explicitly modeling the shadow geometry. Still, they lack control and are prone to visual artifacts. We introduce “Pixel Height”, a novel geometry representation that encodes the correlations between objects, ground, and camera pose. The Pixel Height can be calculated from 3D geometries, manually annotated on 2D images, and it can also be predicted from a single-view RGB image by a supervised approach. It can be used to calculate hard shadows in a 2D image based on the projective geometry, providing precise control of the shadows’ direction and shape. Furthermore, we propose a data-driven soft shadow generator to apply softness to a hard shadow based on a softness input parameter. Qualitative and quantitative evaluations demonstrate that the proposed Pixel Height significantly improves the quality of the shadow generation while allowing for controllability."

"Pioneering dual-encoder pre-training works (e.g., CLIP and ALIGN) have revealed the potential of aligning multi-modal representations with contrastive learning. However, these works require a tremendous amount of data and computational resources (e.g., billion-level web data and hundreds of GPUs), which prevent researchers with limited resources from reproduction and further exploration. To this end, we propose a stack of novel methods, which significantly cut down the heavy resource dependency and allow us to conduct dual-encoder multi-modal representation alignment with limited resources. Besides, we provide a reproducible baseline of competitive results, namely ZeroVL, with only 14M publicly accessible academic datasets and 8 V100 GPUs. Additionally, we collect 100M web data for pre-training, and achieve comparable or superior results than state-of-the-art methods, further proving the effectiveness of our methods on large-scale data. We hope that this work will provide useful data points and experience for future research in contrastive vision-language pre-training. Code is available at https://github.com/zerovl/ZeroVL."

"Recent unsupervised domain adaptation methods have utilized vicinal space between the source and target domains. However, the equilibrium collapse of labels, a problem where the source labels are dominant over the target labels in the predictions of vicinal instances, has never been addressed. In this paper, we propose an instance-wise minimax strategy that minimizes the entropy of high uncertainty instances in the vicinal space to tackle the stated problem. We divide the vicinal space into two subspaces through the solution of the minimax problem: contrastive space and consensus space. In the contrastive space, inter-domain discrepancy is mitigated by constraining instances to have contrastive views and labels, and the consensus space reduces the confusion between intra-domain categories. The effectiveness of our method is demonstrated on public benchmarks, including Office-31, Office-Home, and VisDA-C, achieving state-of-the-art performances. We further show that our method outperforms the current state-of-the-art methods on PACS, which indicates that our instance-wise approach works well for multi-source domain adaptation as well. Code is available at https://github.com/NaJaeMin92/CoVi."

"Unsupervised few-shot learning aims to learn the inductive bias from unlabeled dataset for solving the novel few-shot tasks. The existing unsupervised few-shot learning models and the contrastive learning models follow a unified paradigm. Therefore, we conduct empirical study under this paradigm and find that pairwise contrast, meta losses and large batch size are the important design factors. This results in our CPN (Contrastive Prototypical Network) model, which combines the prototypical loss with pairwise contrast and outperforms the existing models from this paradigm with modestly large batch size. Furthermore, the one-hot prediction target in CPN could lead to learning the sample-specific information. To this end, we propose Wasserstein Confidence Penalty which can impose appropriate penalty on overconfident predictions based on the semantic relationships among pseudo classes. Our full model, CPNWCP (Contrastive Prototypical Network with Wasserstein Confidence Penalty), achieves state-of-the-art performance on miniImageNet and tieredImageNet under unsupervised setting. Our code is available at https://github.com/Haoqing-Wang/CPNWCP."

"Recent contrastive based 3D action representation learning has made great progress. However, the strict positive/negative constraint is yet to be relaxed and the use of non-self positive is yet to be explored. In this paper, a Contrastive Positive Mining (CPM) framework is proposed for unsupervised skeleton 3D action representation learning. The CPM identifies non-self positives in a contextual queue to boost learning. Specifically, the siamese encoders are adopted and trained to match the similarity distributions of the augmented instances in reference to all instances in the contextual queue. By identifying the non-self positive instances in the queue, a positive-enhanced learning strategy is proposed to leverage the knowledge of mined positives to boost the robustness of the learned latent space against intra-class and inter-class diversity. Experimental results have shown that the proposed CPM is effective and outperforms the existing state-of-the-art unsupervised methods on the challenging NTU and PKU-MMD datasets."

"Continuous one-to-many mapping is a less investigated yet important task in both low-level visions and neural image translation. In this paper, we present a new formulation called MonoPix, an unsupervised and contrastive continuous modulation model, and take a step further to enable a pixel-level spatial control which is critical but can not be properly handled previously. The key feature of this work is to model the monotonicity between controlling signals and the domain discriminator with a novel contrastive modulation framework and corresponding monotonicity constraints. We have also introduced a selective inference strategy with logarithmic approximation complexity and support fast domain adaptations. The state-of-the-art performance is validated on a variety of continuous mapping tasks, including AFHQ cat-dog and Yosemite summer-winter translation. The introduced approach also helps to provide a new solution for many low-level tasks like low-light enhancement and natural noise generation, which is beyond the long-established practice of one-to-one training and inference. Code is available at https://github.com/lukun199/MonoPix."

"We introduce a new method for diverse foreground generation with explicit control over various factors. Existing image inpainting based foreground generation methods often struggle to generate diverse results and rarely allow users to explicitly control specific factors of variation (e.g., varying the facial identity or expression for face inpainting results). We leverage contrastive learning with latent codes to generate diverse foreground results for the same masked input. Specifically, we define two sets of latent codes, where one controls a pre-defined factor (“known”), and the other controls the remaining factors (“unknown”). The sampled latent codes from the two sets jointly bi-modulate the convolution kernels to guide the generator to synthesize diverse results. Experiments demonstrate the superiority of our method over state-of-the-arts in result diversity and generation controllability."

"The success of deep learning is usually accompanied by the growth in neural network depth. However, the traditional training method only supervises the neural network at its last layer and propagates the supervision layer-by-layer, which leads to hardship in optimizing the intermediate layers. Recently, deep supervision has been proposed to add auxiliary classifiers to the intermediate layers of deep neural networks. By optimizing these auxiliary classifiers with the supervised task loss, the supervision can be applied to the shallow layers directly. However, deep supervision conflicts with the well-known observation that the shallow layers learn low-level features instead of task-biased high-level semantic features. To address this issue, this paper proposes a novel training framework named Contrastive Deep Supervision, which supervises the intermediate layers with augmentation-based contrastive learning. Experimental results on nine popular datasets with eleven models demonstrate its effects on general image classification, fine-grained image classification and object detection in supervised learning, semi-supervised learning and knowledge distillation. Codes have been released in Github."

"The standard approach to contrastive learning is to maximize the agreement between different views of the data. The views are ordered in pairs, such that they are either positive, encoding different views of the same object, or negative, corresponding to views of different objects. The supervisory signal comes from maximizing the total similarity over positive pairs, while the negative pairs are needed to avoid collapse. In this work, we note that the approach of considering individual pairs cannot account for both intra-set and inter-set similarities when the sets are formed from the views of the data. It thus limits the information content of the supervisory signal available to train representations. We propose to go beyond contrasting individual pairs of objects by focusing on contrasting objects as sets. For this, we use combinatorial quadratic assignment theory designed to evaluate set and graph similarities and derive set-contrastive objective as a regularizer for contrastive learning methods. We conduct experiments and demonstrate that our method improves learned representations for the tasks of metric learning and self-supervised classification."

"Video-based remote physiological measurement utilizes face videos to measure the blood volume change signal, which is also called remote photoplethysmography (rPPG). Supervised methods for rPPG measurements achieve state-of-the-art performance. However, supervised rPPG methods require face videos and ground truth physiological signals for model training. In this paper, we propose an unsupervised rPPG measurement method that does not require ground truth signals for training. We use a 3DCNN model to generate multiple rPPG signals from each video in different spatiotemporal locations and train the model with a contrastive loss where rPPG signals from the same video are pulled together while those from different videos are pushed away. We test on five public datasets, including RGB videos and NIR videos. The results show that our method outperforms the previous unsupervised baseline and achieves accuracies very close to the current best supervised rPPG methods on all five datasets. Furthermore, we also demonstrate that our approach can run at a much faster speed and is more robust to noises than the previous unsupervised baseline. Our code is available at https://github.com/zhaodongsun/contrast-phys."

"Due to their inference, data representation and reconstruction properties, Variational Autoencoders (VAE) have been successfully used in continual learning classification tasks. However, their ability to generate images with specifications corresponding to the classes and databases learned during Continual Learning (CL) is not well understood and catastrophic forgetting remains a significant challenge. In this paper, we firstly analyze the forgetting behaviour of VAEs by developing a new theoretical framework that formulates CL as a dynamic optimal transport problem. This framework proves approximate bounds to the data likelihood without requiring the task information and explains how the prior knowledge is lost during the training process. We then propose a novel memory buffering approach, namely the Online Cooperative Memorization (OCM) framework, which consists of a Short-Term Memory (STM) that continually stores recent samples to provide future information for the model, and a Long-Term Memory (LTM) aiming to preserve a wide diversity of samples. The proposed OCM transfers certain samples from STM to LTM according to the information diversity selection criterion without requiring any supervised signals. The OCM framework is then combined with a dynamic VAE expansion mixture network for further enhancing its performance."

"Deep networks have been shown to suffer from catastrophic forgetting. In this work, we try to alleviate this phenomenon in the field of continual semantic segmentation (CSS). We observe that two main problems lie in existing arts. First, attention is only paid to designing constraints for encoder (i.e., the backbone of segmentation network) or output probabilities. But we find that forgetting also happens in the decoder head and harms the performance greatly. Second, old and new knowledge are entangled in intermediate features when learning new categories, making existing practices hard to balance between plasticity and stability. On these bases, we propose a framework driven by two novel constraints to address the aforementioned problems. First, a structure preserving loss is applied to the decoder’s output to maintain the discriminative power of old classes from two different granularities in embedding space. Second, a feature projection module is adopted to disentangle the process of preserving old knowledge from learning new classes. Extensive evaluations on VOC2012 and ADE20K datasets show the effectiveness of our approach, which significantly outperforms existing state-of-the-art CSS methods."