Papers

Zero-shot learning (ZSL) aims to recognize novel classes by transferring semantic knowledge from seen classes to unseen ones. Semantic knowledge is learned from attribute descriptions shared between different classes, which are strong prior for localization of object attribute for representing discriminative region features enabling significant visual-semantic interaction. Although few attention-based models have attempted to learn such region features in a single image, the transferability and discriminative attribute localization of visual features are typically neglected. In this paper, we propose an attribute-guided Transformer network to learn the attribute localization for discriminative visual-semantic embedding representations in ZSL, termed TransZero. Specifically, TransZero takes a feature augmentation encoder to alleviate the cross-dataset bias between ImageNet and ZSL benchmarks and improve the transferability of visual features by reducing the entangled relative geometry relationships among region features. To learn locality-augmented visual features, TransZero employs a visual-semantic decoder to localize the most relevant image regions to each attributes from a given image under the guidance of attribute semantic information. Then, the locality-augmented visual features and semantic vectors are used for conducting effective visual-semantic interaction in a visual-semantic embedding network. Extensive experiments show that TransZero achieves a new state-of-the-art on three ZSL benchmarks. The codes are available at: \url{https://github.com/shiming-chen/TransZero}.

poster session 6 @ blue 2, poster session 12 @ blue 2, poster session 6, poster session 12

Running Convolutional Neural Networks (CNNs) on resource strictly restricted IoT devices, e.g. microcontroller units (MCU) and field programmable gate array (FPGA), is an important and challenging topic: Tiny machine learning (TinyML). To meet very limited storage and power requirements of IoT devices, it is essential to reduce the parameters of CNN models as much as possible. In this paper, we aim at developing an efficient tiny model with only hundreds of KB parameters. Toward this end, we firstly design a parameter-efficient tiny architecture by introducing dense linear depthwise block. Then, a novel adaptive scale quantization (ASQ) method is proposed for further quantizing tiny models in aggressive low-bit while retaining the accuracy. With the optimized architecture and 4-bit ASQ, we present a family of ultra-lightweight networks, named EtinyNet, that achieves 57.0% ImageNet top-1 accuracy with an extremely tiny model size of 340KB. When deployed on an off-the-shelf commercial microcontroller for object detection tasks, EtinyNet achieves state-of-the-art 56.4% mAP on Pascal VOC. Furthermore, the experimental results on Xilinx compact FPGA indicate that EtinyNet achieves prominent low power of 620mW, about 5.6x lower than existing FPGA designs.

poster session 6 @ blue 2, poster session 12 @ blue 2, poster session 6, poster session 12

In unconstrained scenarios, face recognition remains challenging, particularly when faces are occluded. Existing methods generalize poorly due to the distribution distortion induced by unpredictable occlusions. To tackle this problem, we propose a hierarchical segmentation-based mask learning strategy for face recognition, enhancing occlusion-robustness by integrating segmentation representations of occlusion into face recognition in the latent space. We present a novel multi-scale segmentation-based mask learning (MSML) network, which consists of a face recognition branch (FRB), an occlusion segmentation branch (OSB), and hierarchical elaborate feature masking (FM) operators. With the guidance of hierarchical segmentation representations of occlusion learned by the OSB, the FM operators can generate multi-scale latent masks to eliminate mistaken responses introduced by occlusions and purify the contaminated facial features at multiple layers. In this way, the proposed MSML network can effectively identify and remove the occlusions from feature representations at multiple levels and aggregate features from visible facial areas. Experiments on face verification and recognition under synthetic or realistic occlusions demonstrate the effectiveness of our method compared to state-of-the-art methods.

poster session 6 @ blue 2, poster session 12 @ blue 2, oral session 6 @ blue 2, poster session 6, poster session 12, oral session 6

Graph neural networks (GNNs) have been used to tackle the few-shot learning (FSL) problem and shown great potentials under the transductive setting. However under the inductive setting, existing GNN based methods are less competitive. This is because they use an instance GNN as a label propagation/classification module, which is jointly meta-learned with a feature embedding network. This design is problematic because the classifier needs to adapt quickly to new tasks while the embedding does not. To overcome this problem, in this paper we propose a novel hybrid GNN (HGNN) model consisting of two GNNs, an instance GNN and a prototype GNN. Instead of label propagation, they act as feature embedding adaptation modules for quick adaptation of the meta-learned feature embedding to new tasks. Importantly they are designed to deal with a fundamental yet often neglected challenge in FSL, that is, with only a handful of shots per class, any few-shot classifier would be sensitive to badly sampled shots which are either outliers or can cause inter-class distribution overlapping. Extensive experiments show that our HGNN obtains new state-of-the-art on three FSL benchmarks. The code and models are available at https://github.com/TianyuanYu/HGNN.

poster session 6 @ blue 2, poster session 12 @ blue 2, poster session 6, poster session 12

Semantic segmentation is an important task for scene understanding in self-driving cars and robotics, which aims to assign dense labels for all pixels in the image. Existing work typically improves semantic segmentation performance by exploring different network architectures on a target dataset. Little attention has been paid to build a unified system by simultaneously learning from multiple datasets due to the inherent distribution shift across different datasets. In this paper, we propose a simple, flexible, and general method for semantic segmentation, termed Cross-Dataset Collaborative Learning (CDCL). Our goal is to train a unified model for improving the performance in each dataset by leveraging information from all the datasets. Specifically, we first introduce a family of Dataset-Aware Blocks (DAB) as the fundamental computing units of the network, which help capture homogeneous convolutional representations and heterogeneous statistics across different datasets. Second, we present a Dataset Alternation Training (DAT) mechanism to facilitate the collaborative optimization procedure. We conduct extensive evaluations on diverse semantic segmentation datasets for autonomous driving. Experiments demonstrate that our method consistently achieves notable improvements over prior single-dataset and cross-dataset training methods without introducing extra FLOPs. Particularly, with the same architecture of PSPNet (ResNet-18), our method outperforms the single-dataset baseline by 5.65\%, 6.57\%, and 5.79\% mIoU on the validation sets of Cityscapes, BDD100K, CamVid, respectively. We also apply CDCL for point cloud 3D semantic segmentation and achieve improved performance, which further validates the superiority and generality of our method. Code and models will be released.

poster session 6 @ blue 2, poster session 12 @ blue 2, poster session 6, poster session 12

Image captioning has emerged as an interesting research field in recent years due to its broad application scenarios. The traditional paradigm of image captioning relies on paired image-caption datasets to train the model in a supervised manner. However, creating such paired datasets for every target language is prohibitively expensive, which hinders the extensibility of captioning technology and deprives a large part of the world population of its benefit. In this work, we present a novel unpaired cross-lingual method to generate image captions without relying on any caption corpus in the source or the target language. Specifically, our method consists of two phases: (1) a cross-lingual auto-encoding process, which utilizing a sentence parallel (bitext) corpus to learn the mapping from the source to the target language in the scene graph encoding space and decode sentences in the target language, and (2) a cross-modal unsupervised feature mapping, which seeks to map the encoded scene graph features from image modality to language modality. We verify the effectiveness of our proposed method on the Chinese image caption generation task. The comparisons against several existing methods demonstrate the effectiveness of our approach.

poster session 6 @ blue 2, poster session 12 @ blue 2, oral session 6 @ blue 2, poster session 6, poster session 12, oral session 6

Fine-grained visual classification (FGVC) which aims at recognizing objects from subcategories is a very challenging task due to the inherently subtle inter-class differences. Most existing works mainly tackle this problem by reusing the backbone network to extract features of detected discriminative regions. However, this strategy inevitably complicates the pipeline and pushes the proposed regions to contain most parts of the objects thus fails to locate the really important parts. Recently, vision transformer (ViT) shows its strong performance in the traditional classification task. The self-attention mechanism of the transformer links every patch token to the classification token. In this work, we first evaluate the effectiveness of the ViT framework in the fine-grained recognition setting. Then motivated by the strength of the attention link can be intuitively considered as an indicator of the importance of tokens, we further propose a novel Part Selection Module that can be applied to most of the transformer architectures where we integrate all raw attention weights of the transformer into an attention map for guiding the network to effectively and accurately select discriminative image patches and compute their relations. A contrastive loss is applied to enlarge the distance between feature representations of confusing classes. We name the augmented transformer-based model TransFG and demonstrate the value of it by conducting experiments on five popular fine-grained benchmarks where we achieve state-of-the-art performance. Qualitative results are presented for better understanding of our model.

poster session 6 @ blue 2, poster session 12 @ blue 2, poster session 6, poster session 12

The recently advanced unsupervised learning approaches use the siamese-like framework to compare two "views" from the same image for learning representations. Making the two views distinctive is a core to guarantee that unsupervised methods can learn meaningful information. However, such frameworks are sometimes fragile on overfitting if the augmentations used for generating two views are not strong enough, causing the over-confident issue on the training data. This drawback hinders the model from learning subtle variance and fine-grained information. To address this, in this work we aim to involve the distance concept on label space in the unsupervised learning and let the model be aware of the degree of similarity between positive or negative pairs through mixing the input space, to further work collaboratively for the input and loss spaces. Despite its conceptual simplicity, we show empirically that with the solution -- Unsupervised image mixtures (Un-Mix), we can learn more robust and generalized representations from the transformed input and corresponding new label space. Extensive experiments are conducted on CIFAR-10, CIFAR-100, STL-10, Tiny ImageNet and standard ImageNet with popular unsupervised methods SimCLR, BYOL, MoCo V1&V2, SwAV, etc. Our proposed image mixture and label assignment strategy can obtain consistent improvement by 1~3% following exactly the same hyperparameters and training procedures of the base methods. Code is publicly available at https://github.com/szq0214/Un-Mix.

poster session 6 @ blue 2, poster session 12 @ blue 2, poster session 6, poster session 12

This paper proposes a novel two-stage hypergraph-based framework, dubbed ADaptive Hypergraph Neural Network (AD-HNN) to estimate multiple human poses from a single image, with a keypoint localization network and an Adaptive-Pose Hypergraph Neural Network (AP-HNN) added onto the former network. For providing better guided representations of AP-HNN, we employ a Semantic Interaction Convolution (SIC) module within the initial localization network to acquire more explicit predictions. Build upon this, we design a novel adaptive hypergraph to represent a human body for capturing high-order semantic relations among different joints. Notably, it can adaptively adjust the relations between joints and seek the most reasonable structure for the variable poses to benefit the keypoint localization. These two stages are combined to be trained in an end-to-end fashion. Unlike traditional Graph Convolutional Networks (GCNs) that are based on a fixed tree structure, AP-HNN can deal with ambiguity in human pose estimation. Experimental results demonstrate that the AD-HNN achieves state-of-the-art performance both on the MS-COCO, MPII and CrowdPose datasets.

poster session 6 @ blue 2, poster session 12 @ blue 2, poster session 6, poster session 12

Existing portrait matting methods either require auxiliary inputs that are costly to obtain or involve multiple stages that are computationally expensive, making them less suitable for real-time applications. In this work, we present a light-weight matting objective decomposition network (MODNet) for portrait matting in real-time with a single input image. The key idea behind our efficient design is by optimizing a series of sub-objectives simultaneously via explicit constraints. In addition, MODNet includes two novel techniques for improving model efficiency and robustness. First, an Efficient Atrous Spatial Pyramid Pooling (e-ASPP) module is introduced to fuse multi-scale features for semantic estimation. Second, a self-supervised sub-objectives consistency (SOC) strategy is proposed to adapt MODNet to real-world data to address the domain shift problem common to trimap-free methods. MODNet is easy to be trained in an end-to-end manner. It is much faster than contemporaneous methods and runs at 67 frames per second on a 1080Ti GPU. Experiments show that MODNet outperforms prior trimap-free methods by a large margin on both Adobe Matting Dataset and a carefully designed photographic portrait matting (PPM-100) benchmark proposed by us. Further, MODNet achieves remarkable results on daily photos and videos.

poster session 6 @ blue 2, poster session 12 @ blue 2, poster session 6, poster session 12

One-class classification (OCC) aims to learn an effective data description to enclose all normal training samples and detect anomalies based on the deviation from the data description. Current state-of-the-art OCC models learn a compact normality description by hyper-sphere minimisation, but they often suffer from overfitting the training data, especially when the training set is small or contaminated with anomalous samples. To address this issue, we introduce the interpolated Gaussian descriptor (IGD) method, a novel OCC model that learns a one-class Gaussian anomaly classifier trained with adversarially interpolated training samples. The Gaussian anomaly classifier differentiates the training samples based on their distance to the Gaussian centre and the standard deviation of these distances, offering the model a discriminability w.r.t. the given samples during training. The adversarial interpolation is enforced to consistently learn a smooth Gaussian descriptor, even when the training data is small or contaminated with anomalous samples. This enables our model to learn the data description based on the representative normal samples rather than fringe or anomalous samples, resulting in significantly improved normality description. In extensive experiments on diverse popular benchmarks, including MNIST, Fashion MNIST, CIFAR10, MVTec AD and two medical datasets, IGD achieves better detection accuracy than current state-of-the-art models. IGD also shows better robustness in problems with small or contaminated training sets.

poster session 6 @ blue 2, poster session 12 @ blue 2, oral session 12 @ blue 2, poster session 6, poster session 12, oral session 12

We present an extremely simple Ultra-Resolution Style Transfer framework, termed URST, to flexibly process arbitrary high-resolution images (e.g., 10000x10000 pixels) style transfer for the first time. Most of the existing state-of-the-art methods would fall short due to massive memory cost and small stroke size when processing ultra-high resolution images. URST completely avoids the memory problem caused by ultra-high resolution images by (1) dividing the image into small patches and (2) performing patch-wise style transfer with a novel Thumbnail Instance Normalization (TIN). Specifically, TIN can extract thumbnail features' normalization statistics and apply them to small patches, ensuring the style consistency among different patches. Overall, the URST framework has three merits compared to prior arts. (1) We divide input image into small patches and adopt TIN, successfully transferring image style with arbitrary high-resolution. (2) Experiments show that our URST surpasses existing SOTA methods on ultra-high resolution images benefiting from the effectiveness of the proposed stroke perceptual loss in enlarging the stroke size. (3) Our URST can be easily plugged into most existing style transfer methods and directly improve their performance even without training. Code is available at https://git.io/URST.

poster session 6 @ blue 2, poster session 12 @ blue 2, poster session 6, poster session 12

We introduce Perceiving Stroke-Semantic Context (PerSec), a new approach to self-supervised representation learning tailored for Scene Text Recognition (STR) task. Considering scene text images carry both visual and semantic properties, we equip our PerSec with dual context perceivers which can contrast and learn latent representations from low-level stroke and high-level semantic contextual spaces simultaneously via hierarchical contrastive learning on unlabeled text image data. Experiments in un- and semi-supervised learning settings on STR benchmarks demonstrate our proposed framework can yield a more robust representation for both CTC-based and attention-based decoders than other contrastive learning methods. To fully investigate the potential of our method, we also collect a dataset of 100 million unlabeled text images, named UTI-100M, covering 5 scenes and 4 languages. By leveraging hundred-million-level unlabeled data, our PerSec shows significant performance improvement when fine-tuning the learned representation on the labeled data. Furthermore, we observe that the representation learned by PerSec presents great generalization, especially under few labeled data scenes.

poster session 6 @ blue 2, poster session 12 @ blue 2, poster session 6, poster session 12

Detection of manipulated face images has attracted a lot of interest recently. Various schemes have been proposed to tackle this challenging problem, where the patch-based approaches are shown to be promising. However, the existing patch-based approaches tend to treat different patches equally, which do not fully exploit the patch discrepancy for effective feature learning. In this paper, we propose a Patch Diffusion (PD) module which can be integrated into the existing face manipulation detection networks to boost the performance. The PD consists of Discrepancy Patch Feature Learning (DPFL) and Attention-Aware Message Passing (AMP). The DPFL effectively learns the patch features by a newly designed Pairwise Patch Loss (PPLoss), which takes both the patch importance and correlations into consideration. The AMP diffuses the patches through attention-aware message passing in a graph network, where the attentions are explicitly computed based on the patch features learnt in DPFL. We integrate our PD module into four recent face manipulation detection networks, and carry out the experiments on four popular datasets. The results demonstrate that our PD module is able to boost the performance of the existing networks for face manipulation detection.

poster session 6 @ blue 2, poster session 12 @ blue 2, oral session 6 @ blue 2, poster session 6, poster session 12, oral session 6

Deepfake has ignited hot research interests in both academia and industry due to its potential security threats. Many countermeasures have been proposed to mitigate such risks. Current Deepfake detection methods achieve superior performances in dealing with low-visual-quality Deepfake media which can be distinguished by the obvious visual artifacts. However, with the development of deep generative models, the realism of Deepfake media has been significantly improved and becomes tough challenging to current detection models. In this paper, we propose a frame inference-based detection framework (FInfer) to solve the problem of high-visual-quality Deepfake detection. Specifically, we first learn the referenced representations of the current and future frames’ faces. Then, the current frames’ facial representations are utilized to predict the future frames’ facial representations by using an autoregressive model. Finally, a representation-prediction loss is devised to maximize the discriminability of real videos and fake videos. We demonstrate the effectiveness of our FInfer framework through information theory analyses. The entropy and mutual information analyses indicate the correlation between the predicted representations and referenced representations in real videos is higher than that of high-visual-quality Deepfake videos. Extensive experiments demonstrate the performance of our method is promising in terms of in-dataset detection performance, detection efficiency, and cross-dataset detection performance in high-visual-quality Deepfake videos.

poster session 6 @ blue 2, poster session 12 @ blue 2, oral session 12 @ blue 2, poster session 6, poster session 12, oral session 12

Multi-person pose estimation methods generally follow top-down and bottom-up paradigms, both of which can be considered as two-stage approaches thus leading to the high computation cost and low efficiency. Towards a compact and efficient pipeline for multi-person pose estimation task, in this paper, we propose to represent the human parts as points and present a novel body representation, which leverages an adaptive point set including the human center and seven human-part related points to represent the human instance in a more fine-grained manner. The novel representation is more capable of capturing the various pose deformation and adaptively factorizes the long-range center-to-joint displacement thus delivers a single-stage differentiable network to more precisely regress multi-person pose, termed as AdaptivePose. For inference, our proposed network eliminates the grouping as well as refinements and only needs a single-step disentangling process to form multi-person pose. Without any bells and whistles, we achieve the best speed-accuracy trade-offs of 67.4% AP / 29.4 fps with DLA-34 and 71.3% AP / 9.1 fps with HRNet-W48 on COCO test-dev dataset.

poster session 6 @ blue 2, poster session 12 @ blue 2, poster session 6, poster session 12

Universal domain adaptation (UniDA) aims to transfer knowledge learned from a labeled source domain to an unlabeled target domain under domain shift and category shift. Without prior category overlap information, it is challenging to simultaneously align the common categories between two domains and separate their respective private categories. Additionally, previous studies utilize the source classifier's prediction to obtain various known labels and one generic "unknown" label of target samples. However, over-reliance on learned classifier knowledge is inevitably biased to source data, ignoring the intrinsic structure of target domain. Therefore, in this paper, we propose a novel two-stage UniDA framework called MATHS based on the principle of \textbf{M}utual ne\textbf{A}rest neighbor con\textbf{T}rast and \textbf{H}ybrid prototype di\textbf{S}crimination. In the first stage, we design an efficient mutual nearest neighbor contrastive learning scheme to achieve feature alignment, which exploits the instance-level affinity relationship to uncover the intrinsic structure of two domains. We introduce a bimodality hypothesis for the maximum discriminative probability distribution to detect the possible target private samples, and present a data-based statistical approach to separate the common and private categories. In the second stage, to obtain more reliable label predictions, we propose an incremental pseudo-classifier for target data only, which is driven by the hybrid representative prototypes. A confidence-guided prototype contrastive loss is designed to optimize the category allocation uncertainty via a self-training mechanism. Extensive experiments on three benchmarks demonstrate that MATHS outperforms previous state-of-the-arts on most UniDA settings.

poster session 6 @ blue 2, poster session 12 @ blue 2, poster session 6, poster session 12

Dynamic convolution has achieved significant gain in performance and computational complexity, thanks to its powerful representation capability given limited filter number/layers. However, SOTA dynamic convolution operators are sensitive to input noises (\eg, Gaussian noise, shot noise, \etc) and lack sufficient spatial contextual information in filter generation. To alleviate this inherent weakness, we propose a lightweight and heterogeneous-structure (\ie, static and dynamic) operator, named \emph{Bi-volution}. On the one hand, Bi-volution is designed as a dual-branch structure to fully leverage complementary properties of static/dynamic convolution, which endows Bi-volution more robust properties and higher performance. On the other hand, the Spatial Augmented Kernel Generation module is proposed to improve the dynamic convolution, realizing the learning of spatial context information with negligible additional computational complexity. Extensive experiments illustrate that the ResNet-50 equipped with Bi-volution achieves a highly competitive boost in performance ($+2.8\%$ top-1 accuracy on ImageNet classification, $+2.4\%$ box AP and $+2.2\%$ mask AP on COCO detection and instance segmentation) while maintaining extremely low FLOPs (\ie, ResNet50$@2.7$ GFLOPs). Furthermore, our Bi-volution shows better robustness than dynamic convolution against various noise and input corruptions.

poster session 6 @ blue 2, poster session 12 @ blue 2, poster session 6, poster session 12

Off-the-shelf single-stage multi-person pose regression methods generally leverage the instance score (i.e., confidence of the instance localization) to indicate the pose quality for selecting the pose candidates. We consider that there are two gaps involved in existing paradigm: 1) The instance score is not well interrelated with the pose regression quality. 2) The instance feature representation, which is used for predicting the instance score, does not explicitly encode the structural pose information to predict the reasonable score that represents pose regression quality. To address the aforementioned issues, we propose to learn the pose regression quality-aware representation. Concretely, for the first gap, instead of using the previous instance confidence label (e.g., discrete {1,0} or Gaussian representation) to denote the position and confidence for person instance, we firstly introduce the Consistent Instance Representation (CIR) that unifies the pose regression quality score of instance and the confidence of background into a pixel-wise score map to calibrates the inconsistency between instance score and pose regression quality. To fill the second gap, we further present the Query Encoding Module (QEM) including the Keypoint Query Encoding (KQE) to encode the positional and semantic information for each keypoint and the Pose Query Encoding (PQE) which explicitly encodes the predicted structural pose information to better fit the Consistent Instance Representation (CIR). By using the proposed components, we significantly alleviate the above gaps. Our method outperforms previous single-stage regression-based even bottom-up methods and achieves the state-of-the-art result of 71.7 AP on MS COCO test-dev set.

poster session 6 @ blue 2, poster session 12 @ blue 2, poster session 6, poster session 12

DEtection TRansformer (DETR) is a recently proposed method that streamlines the detection pipeline and achieves competitive results against two-stage detectors such as Faster-RCNN. The DETR models get rid of complex anchor generation and post-processing procedures thereby making the detection pipeline more intuitive. However, the numerous redundant parameters in transformers make the DETR models computation and storage intensive, which seriously hinder them to be deployed on the resources-constrained devices. In this paper, to obtain a compact end-to-end detection framework, we propose to deeply compress the transformers with low-rank tensor decomposition. The basic idea of the tensor-based compression is to represent the large-scale weight matrix in one network layer with a chain of low-order matrices. Furthermore, we propose a gated multi-head attention (GMHA) module to mitigate the accuracy drop of the tensor-compressed DETR models. In GMHA, each attention head has an independent gate to determine the passed attention value. The redundant attention information can be suppressed by adopting the normalized gates. Lastly, to obtain fully compressed DETR models, a low-bitwidth quantization technique is introduced for further reducing the model storage size. Based on the proposed methods, we can achieve significant parameter and model size reduction while maintaining high detection performance. We conduct extensive experiments on the COCO dataset to validate the effectiveness of our tensor-compressed (tensorized) DETR models. The experimental results show that we can attain 3.7 times full model compression with 482 times feed forward network (FFN) parameter reduction and only 0.6 points accuracy drop.

poster session 6 @ blue 2, poster session 12 @ blue 2, poster session 6, poster session 12

Even though graphic layout generation has attracted growing attention recently, it is still challenging to synthesis realistic and diverse layouts, due to the complicated element relationships and varied element arrangements. In this work, we seek to improve the performance of layout generation by incorporating the concept of regions, which consist of a smaller number of elements and appears like a simple layout, into the generation process. Specifically, we leverage Variational Autoencoder (VAE) as the overall architecture and decompose the decoding process into two stages. The first stage predicts representations for regions, and the second stage fills in the detailed position for each element within the region based on the predicted region representation. Compared to prior studies that merely abstract the layout into a list of elements and generate all the element positions in one go, our approach has at least two advantages. First, by the two-stage decoding, our approach decouples the complex layout generation task into several simple layout generation tasks, which reduces the problem difficulty. Second, the predicted regions can help the model roughly know what the graphic layout looks like and serve as global context to improve the generation of detailed element positions. Qualitative and quantitative experiments demonstrate that our approach significantly outperforms the existing methods, especially on the complex graphic layouts.

poster session 6 @ blue 2, poster session 12 @ blue 2, oral session 12 @ blue 2, poster session 6, poster session 12, oral session 12

We present a novel framework that brings the 3D motion retargeting task from controlled environments to in-the-wild scenarios. In particular, our method is capable of retargeting body motion from a character in a 2D monocular video to a 3D character without using any motion capture system or 3D reconstruction procedure. It is designed to leverage massive online videos for unsupervised training, needless of 3D annotations or motion-body pairing information. The proposed method is built upon two novel canonicalization operations, structure canonicalization and view canonicalization. Trained with the canonicalization operations and the derived regularizations, our method learns to factorize a skeleton sequence into three independent semantic subspaces, i.e., motion, structure, and view angle. The disentangled representation enables motion retargeting from 2D to 3D with high precision. Our method achieves superior performance on motion transfer benchmarks with large body variations and challenging actions. Notably, the canonicalized skeleton sequence could serve as a disentangled and interpretable representation of human motion that benefits action analysis and motion retrieval.

poster session 6 @ blue 2, poster session 12 @ blue 2, poster session 6, poster session 12

Universal domain adaptation (UniDA) aims to transfer the knowledge learned from a labeled source domain to an unlabeled target domain without any constraints on the label sets. However, domain shift and category shift make UniDA extremely challenging, mainly attributed to the requirement of identifying both shared “known” samples and private “unknown” samples. Previous methods barely exploit the intrinsic manifold structure relationship between two domains for feature alignment, and they rely on the softmax-based scores with class competition nature to detect underlying “unknown” samples. Therefore, in this paper, we propose a novel eviden\textbf{T}ial \textbf{N}eighborhood con\textbf{T}rastive learning framework called TNT to address these issues. Specifically, TNT first proposes a new domain alignment principle: semantically consistent samples should be geometrically adjacent to each other, whether within or across domains. From this criterion, a cross-domain multi-sample contrastive loss based on mutual nearest neighbors is designed to achieve common category matching and private category separation. Second, toward accurate “unknown” sample detection, TNT introduces a class competition-free uncertainty score from the perspective of evidential deep learning. Instead of setting a single threshold, TNT learns a category-aware heterogeneous threshold vector to reject diverse “unknown” samples. Extensive experiments on three benchmarks demonstrate that TNT significantly outperforms previous state-of-the-art UniDA methods.

poster session 6 @ blue 2, poster session 12 @ blue 2, poster session 6, poster session 12

Pursuing accurate and robust recognizers has been a long-lasting goal for scene text recognition (STR) researchers. Recently, attention-based methods have demonstrated their effectiveness and achieved impressive results on public benchmarks. The attention mechanism enables models to recognize scene text with severe visual distortions by leveraging contextual information. However, recent studies revealed that the implicit over-reliance of context leads to catastrophic out-of-vocabulary performance. On the contrary to the superior accuracy of the seen text, models are prone to misrecognize unseen text even with good image quality. We propose a novel framework, Context-based contrastive learning (ConCLR), to alleviate this issue. Our proposed method first generates characters with different contexts via simple image concatenation operations and then optimizes contrastive loss on their embeddings. By pulling together clusters of identical characters within various contexts and pushing apart clusters of different characters in embedding space, ConCLR suppresses the side-effect of overfitting to specific contexts and learns a more robust representation. Experiments show that ConCLR significantly improves out-of-vocabulary generalization and achieves state-of-the-art performance on public benchmarks together with attention-based recognizers.

poster session 6 @ blue 2, poster session 12 @ blue 2, poster session 6, poster session 12