WACV 2023 论文列表

In this work, we for the first time present a method for detecting labeling errors in image datasets with semantic segmentation, i.e., pixel-wise class labels. Annotation acquisition for semantic segmentation datasets is time-consuming and requires plenty of human labor. In particular, review processes are time consuming and label errors can easily be overlooked by humans. The consequences are biased benchmarks and in extreme cases also performance degradation of deep neural networks (DNNs) trained on such datasets. DNNs for semantic segmentation yield pixel-wise predictions, which makes detection of labeling errors via uncertainty quantification a complex task. Uncertainty is particularly pronounced at the transitions between connected components of the prediction. By lifting the consideration of uncertainty to the level of predicted components, we enable the usage of DNNs together with component-level uncertainty quantification for the detection of labeling errors. We present a principled approach to benchmarking the task of label error detection by dropping labels from the Cityscapes dataset as well from a dataset extracted from the CARLA driving simulator, where in the latter case we have the labels under control. Our experiments show that our approach is able to detect the vast majority of labeling errors while controlling the number of false label error detections. Furthermore, we apply our method to semantic segmentation datasets frequently used by the computer vision community and present a collection of labeling errors along with sample statistics.

Even after decades of research, dynamic scene background reconstruction and foreground object segmentation are still considered as open problems due to various challenges such as illumination changes, camera movements, or background noise caused by air turbulence or moving trees. We propose in this paper to model the background of a frame sequence as a low dimensional manifold using an autoencoder and compare the reconstructed background provided by this autoencoder with the original image to compute the foreground/background segmentation masks. The main novelty of the proposed model is that the autoencoder is also trained to predict the background noise, which allows to compute for each frame a pixel-dependent threshold to perform the foreground segmentation. Although the proposed model does not use any temporal or motion information, it exceeds the state of the art for unsupervised background subtraction on the CDnet 2014 and LASIESTA datasets, with a significant improvement on videos where the camera is moving. It is also able to perform background reconstruction on some non-video image datasets.

A highly accurate but overconfident model is ill-suited for deployment in critical applications such as healthcare and autonomous driving. The classification outcome should reflect a high uncertainty on ambiguous in-distribution samples that lie close to the decision boundary. The model should also refrain from making overconfident decisions on samples that lie far outside its training distribution, far-out-of-distribution (far-OOD), or on unseen samples from novel classes that lie near its training distribution (near-OOD). This paper proposes an application of counterfactual explanations in fixing an over-confident classifier. Specifically, we propose to fine-tune a given pre-trained classifier using augmentations from a counterfactual explainer (ACE) to fix its uncertainty characteristics while retaining its predictive performance. We perform extensive experiments with detecting far-OOD, near-OOD, and ambiguous samples. Our empirical results show that the revised model have improved uncertainty measures, and its performance is competitive to the state-of-the-art methods.

A long-standing goal in the field of sensory substitution is enabling sound perception for deaf and hard of hearing (DHH) people by visualizing audio content. Different from existing models that translate to hand sign language, between speech and text, or text and images, we target immediate and low-level audio to video translation that applies to generic environment sounds as well as human speech. Since such a substitution is artificial, without labels for supervised learning, our core contribution is to build a mapping from audio to video that learns from unpaired examples via high-level constraints. For speech, we additionally disentangle content from style, such as gender and dialect. Qualitative and quantitative results, including a human study, demonstrate that our unpaired translation approach maintains important audio features in the generated video and that videos of faces and numbers are well suited for visualizing high-dimensional audio features that can be parsed by humans to match and distinguish between sounds and words. Project website: https://chunjinsong.github.io/audioviewer

This work proposes a novel method to generate realistic talking head videos using audio and visual streams. We animate a source image by transferring head motion from a driving video using a dense motion field generated using learnable keypoints. We improve the quality of lip sync using audio as an additional input, helping the network to attend to the mouth region. We use additional priors using face segmentation and face mesh to improve the structure of the reconstructed faces. Finally, we improve the visual quality of the generations by incorporating a carefully designed identity-aware generator module. The identity-aware generator takes the source image and the warped motion features as input to generate a high-quality output with fine-grained details. Our method produces state-of-the-art results and generalizes well to unseen faces, languages, and voices. We comprehensively evaluate our approach using multiple metrics and outperforming the current techniques both qualitative and quantitatively. Our work opens up several applications, including enabling low bandwidth video calls. We release a demo video and additional information at http://cvit.iiit.ac.in/research/projects/cvit-projects/avfr

End-to-end Automatic Speech Recognition (ASR) systems based on neural networks have seen large improvements in recent years. The availability of large scale hand-labeled datasets and sufficient computing resources made it possible to train powerful deep neural networks, reaching very low Word Error Rate (WER) on academic benchmarks. However, despite impressive performance on clean audio samples, a drop of performance is often observed on noisy speech. In this work, we propose to improve the noise robustness of the recently proposed Efficient Conformer Connectionist Temporal Classification (CTC)-based architecture by processing both audio and visual modalities. We improve previous lip reading methods using an Efficient Conformer back-end on top of a ResNet-18 visual front-end and by adding intermediate CTC losses between blocks. We condition intermediate block features on early predictions using Inter CTC residual modules to relax the conditional independence assumption of CTC-based models. We also replace the Efficient Conformer grouped attention by a more efficient and simpler attention mechanism that we call patch attention. We experiment with publicly available Lip Reading Sentences 2 (LRS2) and Lip Reading Sentences 3 (LRS3) datasets. Our experiments show that using audio and visual modalities allows to better recognize speech in the presence of environmental noise and significantly accelerate training, reaching lower WER with 4 times less training steps. Our Audio-Visual Efficient Conformer (AVEC) model achieves state-of-the-art performance, reaching WER of 2.3% and 1.8% on LRS2 and LRS3 test sets. Code and pretrained models are available at https://github.com/burchim/AVEC.

In class-incremental semantic segmentation (CISS), deep learning architectures suffer from the critical problems of catastrophic forgetting and semantic background shift. Although recent works focused on these issues, existing classifier initialization methods do not address the background shift problem and assign the same initialization weights to both background and new foreground class classifiers. We propose to address the background shift with a novel classifier initialization method which employs gradient-based attribution to identify the most relevant weights for new classes from the classifier's weights for the previous background and transfers these weights to the new classifier. This warm-start weight initialization provides a general solution applicable to several CISS methods. Furthermore, it accelerates learning of new classes while mitigating forgetting. Our experiments demonstrate significant improvement in mIoU compared to the state-of-the-art CISS methods on the Pascal-VOC 2012, ADE20K and Cityscapes datasets.

Monocular Depth Estimation (MDE) aims to predict pixel-wise depth given a single RGB image. For both, the convolutional as well as the recent attention-based models, encoder-decoder-based architectures have been found to be useful due to the simultaneous requirement of global context and pixel-level resolution. Typically, a skip connection module is used to fuse the encoder and decoder features, which comprises of feature map concatenation followed by a convolution operation. Inspired by the demonstrated benefits of attention in a multitude of computer vision problems, we propose an attention-based fusion of encoder and decoder features. We pose MDE as a pixel query refinement problem, where coarsest-level encoder features are used to initialize pixel-level queries, which are then refined to higher resolutions by the proposed Skip Attention Module (SAM). We formulate the prediction problem as ordinal regression over the bin centers that discretize the continuous depth range and introduce a Bin Center Predictor (BCP) module that predicts bins at the coarsest level using pixel queries. Apart from the benefit of image adaptive depth binning, the proposed design helps learn improved depth embedding in initial pixel queries via direct supervision from the ground truth. Extensive experiments on the two canonical datasets, NYUV2 and KITTI, show that our architecture outperforms the state-of-the-art by 5.3% and 3.9%, respectively, along with an improved generalization performance by 9.4% on the SUNRGBD dataset.

Deep neural networks (DNNs) have rapidly become a de facto choice to medical image understanding tasks. However, DNNs are notoriously fragile to the class imbalance in image classification. We further point out that such imbalance fragility can be amplified when it comes to more sophisticated tasks such as pathology localization, as imbalances in such problems can have highly complex and often implicit forms of presence. For example, different pathology can have different sizes or colors (w.r.t.the background), different underlying demographic distributions, and in general different difficulty levels to recognize, even in a meticulously curated balanced distribution of training data. In this paper, we propose to use pruning to automatically and adaptively identify hard-to-learn (HTL) training samples, and improve pathology localization by attending them explicitly, during training in supervised, semi-supervised, and weakly-supervised settings. Our main inspiration is drawn from the recent finding that deep classification models have difficult-to-memorize samples and those may be effectively exposed through network pruning - and we extend such observation beyond classification for the first time. We also present interesting demographic analysis which illustrates HTLs ability to capture complex demographic imbalances. Our extensive experiments on the Skin Lesion Localization task in multiple training settings by paying additional attention to HTLs show significant improvement of localization performance by 2-3%.

Multi-object tracking (MOT) is a vital component of intelligent video analytics applications such as surveillance and autonomous driving. The time and storage complexity required to execute deep learning models for visual object tracking hinder their adoption on embedded devices with limited computing power. In this paper, we aim to accelerate MOT by transferring the knowledge from high-level features of a complex network (teacher) to a lightweight network (student) at both training and inference times. The proposed AttTrack framework has three key components: 1) cross-model feature learning to align intermediate representations from the teacher and student models, 2) interleaving the execution of the two models at inference time, and 3) incorporating the updated predictions from the teacher model as prior knowledge to assist the student model. Experiments on pedestrian tracking tasks are conducted on the MOT17 and MOT15 datasets using two different object detection backbones YOLOv5 and DLA34 show that AttTrack can significantly improve student model tracking performance while sacrificing only minor degradation of tracking speed.

Industrial defect detection is commonly addressed with anomaly detection (AD) methods where no or only incomplete data of potentially occurring defects is available. This work discovers previously unknown problems of student-teacher approaches for AD and proposes a solution, where two neural networks are trained to produce the same output for the defect-free training examples. The core assumption of student-teacher networks is that the distance between the outputs of both networks is larger for anomalies since they are absent in training. However, previous methods suffer from the similarity of student and teacher architecture, such that the distance is undesirably small for anomalies. For this reason, we propose asymmetric student-teacher networks (AST). We train a normalizing flow for density estimation as a teacher and a conventional feed-forward network as a student to trigger large distances for anomalies: The bijectivity of the normalizing flow enforces a divergence of teacher outputs for anomalies compared to normal data. Outside the training distribution the student cannot imitate this divergence due to its fundamentally different architecture. Our AST network compensates for wrongly estimated likelihoods by a normalizing flow, which was alternatively used for anomaly detection in previous work. We show that our method produces state-of-the-art results on the two currently most relevant defect detection datasets MVTec AD and MVTec 3D-AD regarding image-level anomaly detection on RGB and 3D data.

Turning the weights to zero when training a neural network helps in reducing the computational complexity at inference. To progressively increase the sparsity ratio in the network without causing sharp weight discontinuities during training, our work combines soft-thresholding and straight-through gradient estimation to update the raw, i.e. non-thresholded, version of zeroed weights. Our method, named ST-3 for straight-through/soft-thresholding/sparse-training, obtains SoA results, both in terms of accuracy/sparsity and accuracy/FLOPS trade-offs, when progressively increasing the sparsity ratio in a single training cycle. In particular, despite its simplicity, ST-3 favorably compares to the most recent methods, adopting differentiable formulations or bio-inspired neuroregeneration principles. This suggests that the key ingredients for effective sparsification primarily lie in the ability to give the weights the freedom to evolve smoothly across the zero state while progressively increasing the sparsity ratio. Source code and weights available at https://github.com/vanderschuea/stthree.

Diffusion-based text-to-image generation models like GLIDE and DALLE-2 have gained wide success recently for their superior performance in turning complex text inputs into images of high quality and wide diversity. In particular, they are proven to be very powerful in creating graphic arts of various formats and styles. Although current models supported specifying style formats like oil painting or pencil drawing, fine-grained style features like color distributions and brush strokes are hard to specify as they are randomly picked from a conditional distribution based on the given text input. Here we propose a novel style guidance method to support generating images using arbitrary style guided by a reference image. The generation method does not require a separate style transfer model to generate desired styles while maintaining image quality in generated content as controlled by the text input. Additionally, the guidance method can be applied without a style reference, denoted as self style guidance, to generate images of more diverse styles. Comprehensive experiments prove that the proposed method remains robust and effective in a wide range of conditions, including diverse graphic art forms, image content types and diffusion models.

Although human action anticipation is a task which is inherently multi-modal, state-of-the-art methods on well known action anticipation datasets leverage this data by applying ensemble methods and averaging scores of uni-modal anticipation networks. In this work we introduce transformer based modality fusion techniques, which unify multi-modal data at an early stage. Our Anticipative Feature Fusion Transformer (AFFT) proves to be superior to popular score fusion approaches and presents state-of-the-art results outperforming previous methods on EpicKitchens-100 and EGTEA Gaze+. Our model is easily extensible and allows for adding new modalities without architectural changes. Consequently, we extracted audio features on EpicKitchens-100 which we add to the set of commonly used features in the community.

We present a new method for the unsupervised detection of geometric anomalies in high-resolution 3D point clouds. In particular, we propose an adaptation of the established student-teacher anomaly detection framework to three dimensions. A student network is trained to match the output of a pretrained teacher network on anomaly-free point clouds. When applied to test data, regression errors between the teacher and the student allow reliable localization of anomalous structures. To construct an expressive teacher network that extracts dense local geometric descriptors, we introduce a novel self-supervised pretraining strategy. The teacher is trained by reconstructing local receptive fields and does not require annotations. Extensive experiments on the comprehensive MVTec 3D Anomaly Detection dataset highlight the effectiveness of our approach, which outperforms the existing methods by a large margin. Ablation studies show that our approach meets the requirements of practical applications regarding performance, runtime, and memory consumption.

We study anomaly clustering, grouping data into coherent clusters of anomaly types. This is different from anomaly detection that aims to divide anomalies from normal data. Unlike object-centered image clustering, anomaly clustering is particularly challenging as anomalous patterns are subtle and local. We present a simple yet effective clustering framework using a patch-based pretrained deep embeddings and off-the-shelf clustering methods. We define a distance function between images, each of which is represented as a bag of embeddings, by the Euclidean distance between weighted averaged embeddings. The weight defines the importance of instances (i.e., patch embeddings) in the bag, which may highlight defective regions. We compute weights in an unsupervised way or in a semi-supervised way when labeled normal data is available. Extensive experimental studies show the effectiveness of the proposed clustering framework along with a novel distance function upon exist-ing multiple instance or deep clustering frameworks. Over-all, our framework achieves 0.451 and 0.674 normalized mutual information scores on MVTec object and texture categories and further improve with a few labeled normal data (0.577, 0.669), far exceeding the baselines (0.244, 0.273) or state-of-the-art deep clustering methods (0.176, 0.277).

Plant diseases severely limits agriculture production, necessitating the high-throughput monitoring of plant leaves. Currently, this is formulated as an automatic disease segmentation task addressed via deep learning frameworks. These deep leaning frameworks trained with leaf image data in a supervised paradigm have few limitations, mainly: (1) training datasets are heavily imbalanced towards healthy leaf images, (2) disease region annotation is labour-intensive and (3) due to the heterogeneity of disease symptoms, these frameworks lacks generalisability. In this paper, we reformulate disease segmentation as an anomaly localisation task. Specifically, we introduce a novel unsupervised framework (AnoLeaf) based on an edge-guided inpainting that optimises the learning of contextual attention on only healthy leaf images. The network utilisation on diseased leaf images results in reconstruction of its healthy counterparts, generating an inpainting error. The contextual attention maps reinforce the inpainting error to effectively localise the disease. Thus, AnoLeaf alleviates the acquisition and annotation of rare disease images. Additional experiments on MVTec anomaly detection dataset further demonstrate its generalisability.

This paper studies 3D dense shape correspondence, a key shape analysis application in computer vision and graphics. We introduce a novel hybrid geometric deep learning-based model that learns geometrically meaningful and discretization-independent features. The proposed framework has a U-Net model as the primary node feature extractor, followed by a successive spectral-based graph convolutional network. To create a diverse set of filters, we use anisotropic wavelet basis filters, being sensitive to both different directions and band-passes. This filter set overcomes the common over-smoothing behavior of conventional graph neural networks. To further improve the model's performance, we add a function that perturbs the feature maps in the last layer ahead of fully connected layers, forcing the network to learn more discriminative features overall. The resulting correspondence maps show state-of-the-art performance on the benchmark datasets based on average geodesic errors and superior robustness to discretization in 3D meshes. Our approach provides new insights and practical solutions to the dense shape correspondence research.

Different from the open set recognition, generalized open set recognition learns the most similar known classes for unseen samples using known classes samples and side information of known classes. It is challenging because hierarchically structured side information is distorted when features are embedded in the Euclidean space in existing literature, which incurs the difficulty of identifying the unseen samples. In this paper, we introduce a side information learning algorithm for generalized open set recognition based on the hyperbolic space to alleviate the distortion and accurately identify the unknown samples. Specifically, we propose a hyperbolic side information learning framework to identify the unseen samples and an ancestor search algorithm to search the most similar ancestor from the taxonomy of selected known classes. Experiments on CUB-200 and AWA 2 datasets show that our method improves the performance of generalized open set recognition by a large margin.