WACV 2023 论文列表

Deep learning models are powerful tools for addressing challenging medical imaging problems. However, for an ever-growing range of applications, interpreting a model's prediction remains non-trivial. Understanding decisions made by black-box algorithms is critical, and assessing their fairness and susceptibility to bias is a key step towards healthcare deployment. In this paper, we propose DBCE (Deformation Based Counterfactual Explainability). We optimise a diffeomorphic transformation that deforms a given input image to change the prediction of the model. This provides anatomically meaningful saliency maps indicating tissue atrophy and expansion, which can be easily interpreted by clinicians. In our test case, DBCE replicates the transition of a patient from healthy control (HC) to Alzheimer's disease (AD). We benchmark DBCE against three commonly used saliency methods. We show that it provides more meaningful saliency maps when applied to one subject and disease-consistent atrophy patterns when used over a larger cohort. In addition, our method fulfils a recent sanity check and is repeatable for different model initialisations in contrast to classical sensitivity-based methods.

Weakly-supervised object detection (WSOD) models attempt to leverage image-level annotations in lieu of accurate but costly-to-obtain object localization labels. This oftentimes leads to substandard object detection and localization at inference time. To tackle this issue, we propose D2DF2WOD, a Dual-Domain Fully-to-Weakly Supervised Object Detection framework that leverages synthetic data, annotated with precise object localization, to supplement a natural image target domain, where only image-level labels are available. In its warm-up domain adaptation stage, the model learns a fully-supervised object detector (FSOD) to improve the precision of the object proposals in the target domain, and at the same time learns target-domain-specific and detection-aware proposal features. In its main WSOD stage, a WSOD model is specifically tuned to the target domain. The feature extractor and the object proposal generator of the WSOD model are built upon the fine-tuned FSOD model. We test D2DF2WOD on five dual-domain image benchmarks. The results show that our method results in consistently improved object detection and localization compared with state-of-the-art methods.

Product dimension is a crucial piece of information enabling customers make better buying decisions. E-commerce websites extract dimension attributes to enable customers filter the search results according to their requirements. The existing methods extract dimension attributes from textual data like title and product description. However, this textual information often exists in an ambiguous, disorganised structure. In comparison, images can be used to extract reliable and consistent dimensional information. With this motivation, we hereby propose two novel architecture to extract dimensional information from product images. The first namely Single-Box Classification Network is designed to classify each text token in the image, one at a time, whereas the second architecture namely Multi-Box Classification Network uses a transformer network to classify all the detected text tokens simultaneously. To attain better performance, the proposed architectures are also fused with statistical inferences derived from the product category which further increased the F1-score of the Single-Box Classification Network by 3.78% and Multi-Box Classification Network by 0.9%. We use distance supervision technique to create a large scale automated dataset for pretraining purpose and notice considerable improvement when the models were pretrained on the large data before finetuning. The proposed model achieves a desirable precision of 91.54% at 89.75% recall and outperforms the other state of the art approaches by 4.76% in F1-score.

Semi-supervised learning has the potential to improve the data-efficiency of training data-hungry deep neural networks, which is especially important for medical image analysis tasks where labeled data is scarce. In this work, we present a simple semi-supervised learning method for lesion segmentation tasks based on the ideas of cut-paste augmentation and consistency regularization. By exploiting the mask information available in the labeled data, we synthesize partially labeled samples from the unlabeled images so that the usual supervised learning objective (e.g., binary cross entropy) can be applied. Additionally, we introduce a background consistency term to regularize the training on the unlabeled background regions of the synthetic images. We empirically verify the effectiveness of the proposed method on two public lesion segmentation datasets, including an eye fundus photograph dataset and a brain CT scan dataset. The experiment results indicate that our method achieves consistent and superior performance over other self-training and consistency-based methods without introducing sophisticated network components.

Cross-view geo-localization aims to estimate the GPS location of a query ground-view image by matching it to images from a reference database of geo-tagged aerial images. To address this challenging problem, recent approaches use panoramic ground-view images to increase the range of visibility. Although appealing, panoramic images are not readily available compared to the videos of limited Field-Of-View (FOV) images. In this paper, we present the first cross-view geo-localization method that works on a sequence of limited FOV images. Our model is trained end-to-end to capture the temporal structure that lies within the frames using the attention-based temporal feature aggregation module. To robustly tackle different sequences length and GPS noises during inference, we propose to use a sequential dropout scheme to simulate variant length sequences. To evaluate the proposed approach in realistic settings, we present a new large-scale dataset containing ground-view sequences along with the corresponding aerial-view images. Extensive experiments and comparisons demonstrate the superiority of the proposed approach compared to several competitive baselines.

Multi-task learning has recently become a promising solution for a comprehensive understanding of complex scenes. With an appropriate design multi-task models can not only be memory-efficient but also favour the exchange of complementary signals across tasks. In this work, we jointly address 2D semantic segmentation, and two geometry-related tasks, namely dense depth, surface normal estimation as well as edge estimation showing their benefit on indoor and outdoor datasets. We propose a novel multi-task learning architecture that exploits pair-wise cross-task exchange through correlation-guided attention and self-attention to enhance the average representation learning for all tasks. We conduct extensive experiments considering three multi-task setups, showing the benefit of our proposal in comparison to competitive baselines in both synthetic and real benchmarks. We also extend our method to the novel multi-task unsupervised domain adaptation setting. Our code is open-source.

The demand of high-resolution video contents has grown over the years. However, the delivery of high-resolution video is constrained by either computational resources required for rendering or network bandwidth for remote transmission. To remedy this limitation, we leverage the eye trackers found alongside existing augmented and virtual reality headsets. We propose the application of video super-resolution (VSR) technique to fuse low-resolution context with regional high-resolution context for resource-constrained consumption of high-resolution content without perceivable drop in quality. Eye trackers provide us the gaze direction of a user, aiding us in the extraction of the regional high-resolution context. As only pixels that falls within the gaze region can be resolved by the human eye, a large amount of the delivered content is redundant as we can't perceive the difference in quality of the region beyond the observed region. To generate a visually pleasing frame from the fusion of high-resolution region and low-resolution region, we study the capability of a deep neural network of transferring the context of the observed region to other regions (low-resolution) of the current and future frames. We label this task a Foveated Video Super-Resolution (FVSR), as we need to super-resolve the low-resolution regions of current and future frames through the fusion of pixels from the gaze region. We propose Cross-Resolution Flow Propagation (CRFP) for FVSR. We train and evaluate CRFP on REDS dataset on the task of 8 times FVSR, i.e. a combination of 8 times VSR and the fusion of foveated region. Departing from the conventional evaluation of per frame quality using SSIM or PSNR, we propose the evaluation of past foveated region, measuring the capability of a model to leverage the noise present in eye trackers during FVSR.

Recent years have witnessed significant progress in the field of few-shot image classification while few-shot 3D point cloud classification still remains under-explored. Real-world 3D point cloud data often suffers from occlusions, noise and deformation, which make the few-shot 3D point cloud classification even more challenging. In this paper, we propose a cross-modality feature fusion network, for few-shot 3D point cloud classification, which aims to recognize an object given only a few labeled samples, and provides better performance even with point cloud data with missing points. More specifically, we train two models in parallel. One is a projection-based model with ResNet-18 as the backbone and the other one is a point-based model with a DGCNN backbone. Moreover, we design a Support-Query Mutual Attention (sqMA) module to fully exploit the correlation between support and query features. Extensive experiments on three datasets, namely ModelNet40, ModelNet40-C and ScanObjectNN, show the effectiveness of our method, and its robustness to missing points. Our proposed method outperforms different state-of-the-art baselines on all datasets. The margin of improvement is even larger on the ScanObjectNN dataset, which is collected from real-world scenes and is more challenging with objects having missing points.

Image-sentence retrieval has attracted extensive research attention in multimedia and computer vision due to its promising application. The key issue lies in jointly learning the visual and textual representation to accurately estimate their similarity. To this end, the mainstream schema adopts an object-word based attention to calculate their relevance scores and refine their interactive representations with the attention features, which, however, neglects the context of the object representation on the inter-object relationship that matches the predicates in sentences. In this paper, we propose a Cross-modal Semantic Enhanced Interaction method, termed CMSEI for image-sentence retrieval, which correlates the intra- and inter-modal semantics between objects and words. In particular, we first design the intra-modal spatial and semantic graphs based reasoning to enhance the semantic representations of objects guided by the explicit relationships of the objects' spatial positions and their scene graph. Then the visual and textual semantic representations are refined jointly via the inter-modal interactive attention and the cross-modal alignment. To correlate the context of objects with the textual context, we further refine the visual semantic representation via the cross-level object-sentence and word-image based interactive attention. Experimental results on seven standard evaluation metrics show that the proposed CMSEI outperforms the state-of-the-art and the alternative approaches on MS-COCO and Flickr30K benchmarks.

In this paper, we propose a novel dual-branch Transformation-Synthesis network (TS-Net), for video motion retargeting. Given one subject video and one driving video, TS-Net can produce a new plausible video with the subject appearance of the subject video and motion pattern of the driving video. TS-Net consists of a warp-based transformation branch and a warp-free synthesis branch. The novel design of dual branches combines the strengths of deformation-grid-based transformation and warp-free generation for better identity preservation and robustness to occlusion in the synthesized videos. A mask-aware similarity module is further introduced to the transformation branch to reduce computational overhead. Experimental results on face and dance datasets show that TS-Net achieves better performance in video motion retargeting than several state-of-the-art models as well as its single-branch variants. Our code is available at https://github.com/nihaomiao/WACV23_TSNet.

Most cross-domain unsupervised Video Anomaly Detection (VAD) works assume that at least few task-relevant target domain training data are available for adaptation from the source to the target domain. However, this requires laborious model-tuning by the end-user who may prefer to have a system that works "out-of-the-box". To address such practical scenarios, we identify a novel target domain (inference-time) VAD task where no target domain training data are available. To this end, we propose a new 'Zero-shot Cross-domain Video Anomaly Detection (zxvad)' framework that includes a future-frame prediction generative model setup. Different from prior future-frame prediction models, our model uses a novel Normalcy Classifier module to learn the features of normal event videos by learning how such features are different "relative" to features in pseudo-abnormal examples. A novel Untrained Convolutional Neural Network based Anomaly Synthesis module crafts these pseudo-abnormal examples by adding foreign objects in normal video frames with no extra training cost. With our novel relative normalcy feature learning strategy, zxvad generalizes and learns to distinguish between normal and abnormal frames in a new target domain without adaptation during inference. Through evaluations on common datasets, we show that zxvad outperforms the state-of-the-art (SOTA), regardless of whether task-relevant (i.e., VAD) source training data are available or not. Lastly, zxvad also beats the SOTA methods in inference-time efficiency metrics including the model size, total parameters, GPU energy consumption, and GMACs.

Shoe tread impressions are one of the most common types of evidence left at crime scenes. However, the utility of such evidence is limited by the lack of databases of footwear prints that cover the large and growing number of distinct shoe models. Moreover, the database is preferred to contain the 3D shape, or depth, of shoe-tread photos so as to allow for extracting shoeprints to match a query (crime-scene) print. We propose to address this gap by leveraging shoe-tread photos collected by online retailers. The core challenge is to predict depth maps for these photos. As they do not have ground-truth 3D shapes allowing for training depth predictors, we exploit synthetic data that does. We develop a method termed ShoeRinsics that learns to predict depth by leveraging a mix of fully supervised synthetic data and unsupervised retail image data. In particular, we find domain adaptation and intrinsic image decomposition techniques effectively mitigate the synthetic-real domain gap and yield significantly better depth prediction. To validate our method, we introduce 2 validation sets consisting of shoe-tread image and print pairs and define a benchmarking protocol to quantify the quality of predicted depth. On this benchmark, ShoeRinsics outperforms existing methods of depth prediction and synthetic-to-real domain adaptation.

With the development of the self-attention mechanism, the Transformer model has demonstrated its outstanding performance in the computer vision domain. However, the massive computation brought from the full attention mechanism became a heavy burden for memory consumption. Sequentially, the limitation of memory consumption hinders the deployment of the Transformer model on the embedded system where the computing resources are limited. To remedy this problem, we propose a novel memory economy attention mechanism named Couplformer, which decouples the attention map into two sub-matrices and generates the alignment scores from spatial information. Our method enables the Transformer model to improve time and memory efficiency while maintaining expressive power. A series of different scale image classification tasks are applied to evaluate the effectiveness of our model. The result of experiments shows that on the ImageNet-1K classification task, the Couplformer can significantly decrease 42% memory consumption compared with the regular Transformer. Meanwhile, it accesses sufficient accuracy requirements, which outperforms 0.56% on Top-1 accuracy and occupies the same memory footprint. Besides, the Couplformer achieves state-of-art performance in MS COCO 2017 object detection and instance segmentation tasks. As a result, the Couplformer can serve as an efficient backbone in visual tasks and provide a novel perspective on deploying attention mechanisms for researchers.

3D scene understanding is an important problem that has experienced great progress in recent years, in large part due to the development of state-of-the-art methods for 3D object detection. However, the performance of 3D object detectors can suffer in scenarios where extreme occlusion of objects is present, or the number of object classes is large. In this paper, we study the problem of inferring 3D counts from densely packed scenes with heterogeneous objects. This problem has applications to important tasks such as inventory management or automatic crop yield estimation. We propose a novel regression-based method, CountNet3D, that uses mature 2D object detectors for finegrained classification and localization, and a PointNet backbone for geometric embedding. The network processes fused data from images and point clouds for end-to-end learning of counts. We perform experiments on a novel synthetic dataset for inventory management in retail, which we construct and make publicly available to the community. Our results show that regression-based 3D counting methods systematically outperform detection-based methods, and reveal that directly learning from raw point clouds greatly assists count estimation under extreme occlusion. Finally, we study the effectiveness of CountNet3D on a large dataset of real-world scenes where extreme occlusion is present and achieve an error rate of 11.01%.

In this work we address multi-target domain adaptation (MTDA) in semantic segmentation, which consists in adapting a single model from an annotated source dataset to multiple unannotated target datasets that differ in their underlying data distributions. To address MTDA, we propose a self-training strategy that employs pseudo-labels to induce cooperation among multiple domain-specific classifiers. We employ feature stylization as an efficient way to generate image views that forms an integral part of self-training. Additionally, to prevent the network from overfitting to noisy pseudo-labels, we devise a rectification strategy that leverages the predictions from different classifiers to estimate the quality of pseudo-labels. Our extensive experiments on numerous settings, based on four different semantic segmentation datasets, validates the effectiveness of the proposed self-training strategy and shows that our method outperforms state-of-the-art MTDA approaches.

3D face modeling has been an active area of research in computer vision and computer graphics, fueling applications ranging from facial expression transfer in virtual avatars to synthetic data generation. Existing 3D deep learning generative models (e.g., VAE, GANs) allow generating compact face representations (both shape and texture) that can model non-linearities in the shape and appearance space (e.g., scatter effects, specularities,..). However, they lack the capability to control the generation of subtle expressions. This paper proposes a new 3D face generative model that can decouple identity and expression and provides granular control over expressions. In particular, we propose using a pair of supervised auto-encoder and generative adversarial networks to produce high-quality 3D faces, both in terms of appearance and shape. Experimental results in the generation of 3D faces learned with holistic expression labels, or Action Unit (AU) labels, show how we can decouple identity and expression; gaining fine-control over expressions while preserving identity.

We present Control-NeRF, a method for performing flexible, 3D-aware image content manipulation while enabling high-quality novel view synthesis, from a set of posed input images. NeRF-based approaches are effective for novel view synthesis, however such models memorize the radiance for every point in a scene within a neural network. Since these models are scene-specific and lack a 3D scene representation, classical editing such as shape manipulation, or combining scenes is not possible. While there are some recent hybrid approaches that combine NeRF with external scene representations such as sparse voxels, planes, hash tables, etc., they focus mostly on efficiency and don't explore the scene editing and manipulation capabilities of hybrid approaches. With the aim of exploring controllable scene representations for novel view synthesis, our model couples learnt scene-specific 3D feature volumes with a general NeRF rendering network. We can generalize to novel scenes by optimizing only the scene-specific 3D feature volume, while keeping the parameters of the rendering network fixed. Since the feature volumes are independent of the rendering model, we can manipulate and combine scenes by editing their corresponding feature volumes. The edited volume can then be plugged into the rendering model to synthesise high-quality novel views. We demonstrate scene manipulations including: scene mixing; applying rigid and non-rigid transformations; inserting, moving and deleting objects in a scene; while producing photo-realistic novel-view synthesis results.

Video summarization aims to select a most informative subset of frames in a video to facilitate efficient video browsing. Unsupervised methods usually rely on heuristic training objectives such as diversity and representativeness. However, such methods need to bootstrap the online-generated summaries to compute the objectives for importance score regression. We consider such a pipeline inefficient and seek to directly quantify the frame-level importance with the help of contrastive losses in the representation learning literature. Leveraging the contrastive losses, we propose three metrics featuring a desirable key frame: local dissimilarity, global consistency, and uniqueness. With features pre-trained on an image classification task, the metrics can already yield high-quality importance scores, demonstrating better or competitive performance compared with past heavily-trained methods. We show that by refining the pre-trained features with contrastive learning, the frame-level importance scores can be further improved, and the model can learn from random videos and generalize to test videos with decent performance.

We consider the problem of image retrieval where query images during testing belong to classes and domains both unseen during training. This requires learning a feature space that has the ability to generalize across both classes and domains. To this end, we propose semantic contrastive concept network (SCNNet), a new learning framework that helps take a step towards class and domain generalization in a principled fashion. Unlike existing methods that rely on global object representations, SCNNet proposes to learn a set of local concept vectors to facilitate unseen-class generalization. To this end, SCNNet's key innovations include (a) a novel trainable local concept extraction module that learns an orthonormal set of basis vectors, and (b) computes local features for any unseen-class data as a linear combination of the learned basis set. Next, to enable unseen-domain generalization, SCNNet proposes to generate supervisory signals from an adjacent data modality, i.e., natural language, by mining freely available textual label information associated with images. SCNNet derives these signals from our novel trainable semantic ordinal distance constraints that ensure semantic consistency between pairs of images sampled from different domains. Both the proposed modules above enable end-to-end training of the SCNNet, resulting in a model that helps establish state-of-the-art performance on the standard DomainNet, PACS, and Sketchy benchmark datasets with average Prec@200 improvements of 42.6%, 6.5%, and 13.6% respectively over the most recently reported results.