WACV 2023 论文列表

This paper looks at semi-supervised learning (SSL) for image-based text recognition. One of the most popular SSL approaches is pseudo-labeling (PL). PL approaches assign labels to unlabeled data before re-training the model with a combination of labeled and pseudo-labeled data. However, PL methods are severely degraded by noise and are prone to over-fitting to noisy labels, due to the inclusion of erroneous high confidence pseudo-labels generated from poorly calibrated models, thus, rendering threshold-based selection ineffective. Moreover, the combinatorial complexity of the hypothesis space and the error accumulation due to multiple incorrect autoregressive steps posit pseudo-labeling challenging for sequential self-training. To this end, we propose a pseudo-label generation and an uncertainty-based data selection framework for semi-supervised text recognition. We first use Beam-Search inference to yield highly probable hypotheses to assign pseudo-labels to the unlabelled examples. Then we adopt an ensemble of models, sampled by applying dropout, to obtain a robust estimate of the uncertainty associated with the prediction, considering both the character-level and word-level predictive distribution to select good quality pseudo-labels. Extensive experiments on several benchmark handwriting and scene-text datasets show that our method outperforms the baseline approaches and the previous state-of-the-art semi-supervised text-recognition methods.

Separating diffuse and specular reflection components observed on an object surface is important for preprocessing of various computer vision techniques. Conventionally, diffuse-specular separation based on the polarimetric and color clues assumes that the diffuse/specular reflection components are unpolarized/partially polarized under unpolarized light sources. However, the diffuse reflection component is partially polarized in fact, because the diffuse reflectance is maximal when the polarization direction is parallel to the outgoing plane. Accordingly, we propose a method for separating partially-polarized diffuse and specular reflection components on the basis of the polarization reflection model and the dichromatic reflection model. In particular, our method enables us not only to achieve diffuse-specular separation but also to estimate the polarimetric properties of the object surface from a single color polarization image. We experimentally confirmed that our method performs better than the method assuming unpolarized diffuse reflection components.

Labeled data is a critical resource for training and evaluating machine learning models. However, many real-life datasets are only partially labeled. We propose a semi-supervised machine learning training strategy to improve event detection performance on sequential data, such as video recordings, when only sparse labels are available, such as event start times without their corresponding end times. Our method uses noisy guesses of the events' end times to train event detection models. Depending on how conservative these guesses are, mislabeled samples may be introduced into the training set. We further propose a mathematical model for explaining and estimating the evolution of the classification performance for increasingly noisier end time estimates. We show that neural networks can improve their detection performance by leveraging more training data with less conservative approximations despite the higher proportion of incorrect labels. We adapt sequential versions of CIFAR-10 and MNIST, and use the Berkeley MHAD and HMBD51 video datasets to empirically evaluate our method, and find that our risk-tolerant strategy outperforms conservative estimates by 3.5 points of mean average precision for CIFAR, 30 points for MNIST, 3 points for MHAD, and 14 points for HMBD51. Then, we leverage the proposed training strategy to tackle a real-life application: processing continuous video recordings of epilepsy patients, and show that our method outperforms baseline labeling methods by 17 points of average precision, and reaches a classification performance similar to that of fully supervised models.

Convolutional neural networks have been successful in restoring images captured under poor illumination conditions addressing multiple challenges such as contrast enhancement, denoising, and color cast removal. Nevertheless, such approaches require a large number of paired low-light and ground truth images for training. Thus, we study the problem of semi-supervised learning for low-light image restoration when limited low-light images have ground truth labels. Our main contributions in this work are twofold. We first deploy an ensemble of low-light restoration networks to restore the unlabeled images and generate a set of potential pseudo-labels. We model the contrast distortions in the labeled set to generate different sets of training data and create the ensemble of networks. We then design a contrastive self-supervised learning based image quality measure to obtain the pseudo-label among the images restored by the ensemble. We show that training the restoration network with the pseudo-labels allows us to achieve excellent restoration performance even with very few labeled pairs. We conduct extensive experiments on three popular low-light image restoration datasets to show the superior performance of our semi-supervised low-light image restoration compared to other approaches.

We present a new semi-supervised domain adaptation framework that combines a novel auto-encoder-based domain adaptation model with a simultaneous learning scheme providing stable improvements over state-of-the-art domain adaptation models. Our framework holds strong distribution matching property by training both source and target auto-encoders using a novel simultaneous learning scheme on a single graph with an optimally modified MMD loss objective function. Additionally, we design a semi-supervised classification approach by transferring the aligned domain invariant feature spaces from source domain to the target domain. We evaluate on three datasets and show proof that our framework can effectively solve both fragile convergence (adversarial) and weak distribution matching problems between source and target feature space (discrepancy) with a high 'speed' of adaptation requiring a very low number of iterations.

Multi-task learning (MTL) paradigm focuses on jointly learning two or more tasks, aiming for an improvement w.r.t model's generalizability, performance, and training/inference memory footprint. The aforementioned benefits become ever so indispensable in the case of training for vision-related dense prediction tasks. In this work, we tackle the MTL problem of two dense tasks, i.e., semantic segmentation and depth estimation, and present a novel attention module called Cross-Channel Attention Module (CCAM), which facilitates effective feature sharing along each channel between the two tasks, leading to mutual performance gain with a negligible increase in trainable parameters. In a symbiotic spirit, we also formulate novel data augmentations for the semantic segmentation task using predicted depth called AffineMix, and one using predicted semantics called ColorAug, for depth estimation task. Finally, we validate the performance gain of the proposed method on the Cityscapes and ScanNet dataset. which helps us achieve state-of-the-art results for a semi-supervised joint model based on depth estimation and semantic segmentation.

Zero-shot temporal activity detection (ZSTAD) is the problem of simultaneous temporal localization and classification of activity segments that are previously unseen during training. This is achieved by transferring the knowledge learned from semantically-related seen activities. This ability to reason about unseen concepts without supervision makes ZSTAD very promising for applications where the acquisition of annotated training videos is difficult. In this paper, we design a transformer-based framework titled TranZAD, which streamlines the detection of unseen activities by casting ZSTAD as a direct set-prediction problem, removing the need for hand-crafted designs and manual post-processing. We show how a semantic information-guided contrastive learning strategy can effectively train TranZAD for the zero-shot setting, enabling the efficient transfer of knowledge from the seen to the unseen activities. To reduce confusion between unseen activities and unrelated background information in videos, we introduce a more efficient method of computing the background class embedding by dynamically adapting it as part of the end-to-end learning. Additionally, unlike existing work on ZSTAD, we do not assume the knowledge of which classes are unseen during training and use the visual and semantic information of only the seen classes for the knowledge transfer. This makes TranZAD more viable for practical scenarios, which we evaluate by conducting extensive experiments on Thumos'14 and Charades.

Semantic segmentation of degraded images is important for practical applications such as autonomous driving and surveillance systems. The degradation level, which represents the strength of degradation, is usually unknown in practice. Therefore, the semantic segmentation algorithm needs to take account of various levels of degradation. In this paper, we propose a convolutional neural network of semantic segmentation which can cope with various levels of degradation. The proposed network is based on the knowledge distillation from a source network trained with only clean images. More concretely, the proposed network is trained to acquire multi-layer features keeping consistency with the source network, while adjusting for various levels of degradation. The effectiveness of the proposed method is confirmed for different types of degradations: JPEG distortion, Gaussian blur and salt&pepper noise. The experimental comparisons validate that the proposed network outperforms existing networks for semantic segmentation of degraded images with various degradation levels.

Semantic segmentation in large-scale aerial images is an extremely challenging task. On one hand, the limited ground truth, as compared to the vast area the images cover, greatly hinders the development of supervised representation learning. On the other hand, the large footprint from remote sensing raises new challenges for semantic segmentation. In addition, the complex and ever changing image acquisition conditions further complicate the problem where domain shifting commonly occurs. In this paper, we exploit self-supervised contrastive learning (CL) methodologies for semantic segmentation in aerial imagery. In addition to performing CL at the feature level as most practices do, we add another level of contrastive learning, at the semantic level, taking advantage of the segmentation output from the downstream task. Further, we embed local mutual information in the semantic-level CL to enforce local consistency. This has largely enhanced the representation power at each pixel and improved the generalization capacity of the trained model. We refer to the proposed approach as multi-level contrastive learning with local consistency (mCL-LC). The experimental results on different benchmarks indicate that the proposed mCL-LC exhibits superior performance as compared to other state-of-the-art contrastive learning frameworks for the semantic segmentation task. mCL-LC also carries better generalization capacity especially when domain shifting exists.

Using deep learning, we now have the ability to create exceptionally good semantic segmentation systems; however, collecting the prerequisite pixel-wise annotations for training images remains expensive and time-consuming. Therefore, it would be ideal to minimize the number of human annotations needed when creating a new dataset. Here, we address this problem by proposing a novel algorithm that combines active learning and semi-supervised learning. Active learning is an approach for identifying the best unlabeled samples to annotate. While there has been work on active learning for segmentation, most methods require annotating all pixel objects in each image, rather than only the most informative regions. We argue that this is inefficient. Instead, our active learning approach aims to minimize the number of annotations per image. Our method is enriched with semi-supervised learning, where we use pseudo labels generated with a teacher-student framework to identify image regions that help disambiguate confused classes. We also integrate mechanisms that enable better performance on imbalanced label distributions, which have not been studied previously for active learning in semantic segmentation. In experiments on the CamVid and CityScapes datasets, our method obtains over 95% of the network's performance on the full-training set using less than 17% of the training data, whereas the previous state of the art required 40% of the training data.

The ability of few-shot learning (FSL) is a basic requirement of intelligent agent learning in the open visual world. However, existing deep learning systems rely too heavily on large numbers of training samples, making it hard to learn new categories efficiently from limited size of training data. Two key challenges of FSL are insufficient comprehension and imperfect modeling of the few-shot novel class. For insufficient visual comprehension, semantic knowledge which is information from other modalities can help replenish the understanding of novel classes. But even so, most works still suffer from the second challenge because the single global class prototype they adopted is extremely unstable and imperfect given the larger intra-class variation and harder inter-class discrimination in FSL scenario. Thus, we propose to represent each class by its several different parts with the help of class semantic knowledge. Since we can never pre-define parts for unknown novel classes, we embed them in a latent manner. Concretely, we train a generator that takes the class semantic knowledge as input and outputs several filters of class-specific semantic latent parts. By applying each part filter, our model can pay attention to corresponding local regions containing each part. At the inference stage, the classification is conducted by comparing the similarities between those parts. Experiments on several FSL benchmarks demonstrate the effectiveness of our proposed method and show its potential to go beyond class recognition to class understanding. Furthermore, we also find when semantic knowledge is more visualized and customized, it will be more helpful in the FSL task.

We propose a self-supervised method for relative pose estimation for road scenes. By exploiting the approximate planarity of the local ground plane, we can extract a self-supervision signal via cross-projection between images using a homography derived from estimated ground-relative pose. We augment cross-projected perceptual loss by including classical image alignment in the network training loop. We use pretrained semantic segmentation and optical flow to extract ground plane correspondences between approximately aligned images and RANSAC to find the best fitting homography. By decomposing to ground-relative pose, we obtain pseudo labels that can be used for direct supervision. We show that this extremely simple geometric model is competitive for visual odometry with much more complex self-supervised methods that must learn depth estimation in conjunction with relative pose. Code and result videos: github.com/brucemuller/homographyVO.

While self-supervised learning has been shown to benefit a number of vision tasks, existing techniques mainly focus on image-level manipulation, which may not generalize well to downstream tasks at patch or pixel levels. Moreover, existing SSL methods might not sufficiently describe and associate the above representations within and across image scales. In this paper, we propose a Self-Supervised Pyramid Representation Learning (SS-PRL) framework. The proposed SS-PRL is designed to derive pyramid representations at patch levels via learning proper prototypes, with additional learners to observe and relate inherent semantic information within an image. In particular, we present a cross-scale patch-level correlation learning in SS-PRL, which allows the model to aggregate and associate information learned across patch scales. We show that, with our proposed SS-PRL for model pre-training, one can easily adapt and fine-tune the models for a variety of applications including multi-label classification, object detection, and instance segmentation.

Self-supervised monocular depth estimation (MDE) models universally suffer from the notorious edge-fattening issue. Triplet loss, popular for metric learning, has made a great success in many computer vision tasks. In this paper, we redesign the patch-based triplet loss in MDE to alleviate the ubiquitous edge-fattening issue. We show two drawbacks of the raw triplet loss in MDE and demonstrate our problem-driven redesigns. First, we present a min. operator based strategy applied to all negative samples, to prevent well-performing negatives sheltering the error of edge-fattening negatives. Second, we split the anchor-positive distance and anchor-negative distance from within the original triplet, which directly optimizes the positives without any mutual effect with the negatives. Extensive experiments show the combination of these two small redesigns can achieve unprecedented results: Our powerful and versatile triplet loss not only makes our model outperform all previous SoTA by a large margin, but also provides substantial performance boosts to a large number of existing models, while introducing no extra inference computation at all.

Recently, thermal image based 3D understanding is gradually attracting attention for an illumination condition agnostic machine vision. However, the difficulty of the thermal image lies in insufficient training supervision due to its low-contrast and textureless properties. Also, introducing additional modality requires further constraints such as complicated multi-sensor calibration and synchronized data acquisition. To leverage additional modality information without such constraints, we propose a novel training framework that consists of self-supervised learning of unpaired multi-spectral images and feature-level adversarial adaptation. In the training stage, we utilize unpaired RGB/thermal video and partially shared network architecture consisting of modality-specific feature extractors and modality-independent decoder. Through the shared network design, the depth decoder can leverage the self-supervised signal of the unpaired RGB images. Feature-level adversarial adaptation minimizes the gap between RGB and thermal features and eventually makes the thermal encoder extract representative and informative features. Based on the proposed method, the trained depth network shows outperformed results than previous state-of-the-art methods.

The computer-assisted radiologic informative report is currently emerging in dental practice to facilitate dental care and reduce time consumption in manual panoramic radiographic interpretation. However, the amount of dental radiographs for training is very limited, particularly from the point of view of deep learning. This study aims to utilize recent self-supervised learning methods like SimMIM and UM-MAE to increase the model efficiency and understanding of the limited number of dental radiographs. We use the Swin Transformer for teeth numbering, detection of dental restorations, and instance segmentation tasks. To the best of our knowledge, this is the first study that applied self-supervised learning methods to Swin Transformer on dental panoramic radiographs. Our results show that the SimMIM method obtained the highest performance of 90.4% and 88.9% on detecting teeth and dental restorations and instance segmentation, respectively, increasing the average precision by 13.4 and 12.8 over the random initialization baseline. Moreover, we augment and correct the existing dataset of panoramic radiographs. The code and the dataset are available at https://github.com/AmaniHAlmalki/DentalMIM.

We present a self-supervised learning (SSL) method suitable for semi-global tasks such as object detection and semantic segmentation. We enforce local consistency between self-learned features, representing corresponding image locations of transformed versions of the same image, by minimizing a pixel-level local contrastive (LC) loss during training. LC-loss can be added to existing self-supervised learning methods with minimal overhead. We evaluate our SSL approach on two downstream tasks -- object detection and semantic segmentation, using COCO, PASCAL VOC, and CityScapes datasets. Our method outperforms the existing state-of-the-art SSL approaches by 1.9% on COCO object detection, 1.4% on PASCAL VOC detection, and 0.6% on CityScapes segmentation.

Rapid dissemination of misinformation is a major societal problem receiving increasing attention. Unlike Deepfake, Out-of-Context misinformation, in which the unaltered unimode contents (e.g. image, text) of a multi-modal news sample are combined in an out-of-context manner to generate deception, requires limited technical expertise to create. Therefore, it is more prevalent a means to confuse readers. Most existing approaches extract features from its uni-mode counterparts to concatenate and train a model for the misinformation classification task. In this paper, we design a self-supervised feature representation learning strategy that aims to attain the multi-task objectives: (1) task-agnostic, which evaluates the intra- and inter-mode representational consistencies for improved alignments across related models; (2) task-specific, which estimates the category-specific multi-modal knowledge to enable the classifier to derive more discriminative predictive distributions. To compensate for the dearth of annotated data representing varied types of misinformation, the proposed Self-Supervised Distilled Learner (SSDL) utilizes a Teacher network to weakly guide a Student network to mimic a similar decision pattern as the teacher. The two-phased learning of SSDL can be summarized as: initial pretraining of the Student model using a combination of contrastive self-supervised task-agnostic objective and supervised task-specific adjustment in parallel; finetuning the Student model via self-supervised knowledge distillation blended with the supervised objective of decision alignment. In addition to the consistent out-performances over the existing baselines that demonstrate the feasibility of our approach, the explainability capacity of the proposed SSDL also helps users visualize the reasoning behind a specific prediction made by the model.

Video provides us with the spatio-temporal consistency needed for visual learning. Recent approaches have utilized this signal to learn correspondence estimation from close-by frame pairs. However, by only relying on close-by frame pairs, those approaches miss out on the richer long-range consistency between more distant overlapping frames. To address this, we propose a self-supervised approach for correspondence estimation that learns from multiview consistency in short RGB-D video sequences. Our approach combines pairwise correspondence estimation and registration with a novel SE(3) transformation synchronization algorithm. Our key insight is that self-supervised multiview registration allows us to obtain correspondences over longer time frames, which increases both the diversity and difficulty of sampled pairs. We evaluate our approach on indoor scenes for correspondence estimation and RGB-D pointcloud registration and find that we can perform on-par with prior supervised approaches.