We introduce a novel unsupervised domain adaptation approach for object detection. aim to alleviate the imperfect translation problem of pixel-level adaptations, and source-biased discriminativity feature-level adaptations simultaneously. Our is composed two stages, i.e., Domain Diversification (DD) Multi-domain-invariant Representation Learning (MRL). At DD stage, we diversify distribution labeled data by generating various distinctive shifted domains from source domain. MRL apply...

Visible-infrared person re-identification (VI-ReID) is an important task in night-time surveillance applications, since visible cameras are difficult to capture valid appearance information under poor illumination conditions. Compared traditional that handles only the intra-modality discrepancy, VI-ReID suffers from additional cross-modality discrepancy caused by different types of imaging systems. To reduce both intra- and discrepancies, we propose a Hierarchical Cross-Modality...

Deep learning-based semantic segmentation methods have an intrinsic limitation that training a model requires large amount of data with pixel-level annotations. To address this challenging issue, many researchers give attention to unsupervised domain adaptation for segmentation. Unsupervised seeks adapt the trained on source target domain. In paper, we introduce self-ensembling technique, one successful in classification. However, applying is very difficult because heavily-tuned manual...

Deep learning-based object detectors have shown remarkable improvements. However, supervised methods perform poorly when the train data and test different distributions. To address issue, domain adaptation transfers knowledge from label-sufficient (source domain) to label-scarce (target domain). Self-training is one of powerful ways achieve since it helps class-wise adaptation. Unfortunately, a naive approach that utilizes pseudo-labels as ground-truth degenerates performance due incorrect...

Although supervised person re-identification (Re-ID) methods have shown impressive performance, they suffer from a poor generalization capability on unseen domains. Therefore, generalizable Re-ID has recently attracted growing attention. Many existing employed an instance normalization technique to reduce style variations, but the loss of discriminative information could not be avoided. In this paper, we propose novel framework, named Meta Batch-Instance Normalization (MetaBIN). Our main...

Thesedays, Convolutional Neural Networks are widely used in semantic segmentation. However, since CNN-based segmentation networks produce low-resolution outputs with rich information, it is inevitable that spatial details (e.g., small objects and fine boundary information) of results will be lost. To address this problem, motivated by a variational approach to image (i.e., level set theory), we propose novel loss function called the which designed refine results. deal multiple classes an...

To learn target discriminative representations, using pseudo-labels is a simple yet effective approach for unsupervised domain adaptation. However, the existence of false pseudo-labels, which may have detrimental influence on learning remains major challenge. overcome this issue, we propose pseudo-labeling curriculum based density-based clustering algorithm. Since samples with high density values are more likely to correct leverage these subsets train our network at early stage, and utilize...

We present a comparative study on how and why contrastive learning (CL) masked image modeling (MIM) differ in their representations performance of downstream tasks. In particular, we demonstrate that self-supervised Vision Transformers (ViTs) have the following properties: (1) CL trains self-attentions to capture longer-range global patterns than MIM, such as shape an object, especially later layers ViT architecture. This property helps ViTs linearly separate images representation spaces....

Bootstrapping, which enables the full homomorphic encryption scheme that can perform an infinite number of operations by restoring modulus ciphertext with a small modulus, is essential step in encryption. However, bootstrapping most time and memory consuming all operations. As we increase precision bootstrapping, large amount computational resources required. Specifically, for any previous bootstrap designs, limited rescaling precision.

As anomaly detection for electrical power steering (EPS) systems has been centralized using model- and knowledge-based approaches, EPS system have become complex more sophisticated, thereby requiring enhanced reliability safety. Since most current methods rely on prior knowledge, it is difficult to identify new or previously unknown anomalies. In this paper, we propose a deep learning approach that consists of two-stage process an autoencoder long short-term memory (LSTM) detect anomalies in...

We introduce a novel unsupervised domain adaptation approach for object detection. aim to alleviate the imperfect translation problem of pixel-level adaptations, and source-biased discriminativity feature-level adaptations simultaneously. Our is composed two stages, i.e., Domain Diversification (DD) Multi-domain-invariant Representation Learning (MRL). At DD stage, we diversify distribution labeled data by generating various distinctive shifted domains from source domain. MRL apply...

Deep learning-based object detectors have shown remarkable improvements. However, supervised methods perform poorly when the train data and test different distributions. To address issue, domain adaptation transfers knowledge from label-sufficient (source domain) to label-scarce (target domain). Self-training is one of powerful ways achieve since it helps class-wise adaptation. Unfortunately, a naive approach that utilizes pseudo-labels as ground-truth degenerates performance due incorrect...

Although unsupervised domain adaptation methods have been widely adopted across several computer vision tasks, it is more desirable if we can exploit a few labeled data from new domains encountered in real application. The novel setting of the semi-supervised (SSDA) problem shares challenges with and learning problem. However, recent study shows that conventional often result less effective or negative transfer SSDA In order to interpret observation address problem, this paper, raise...

Deep learning-based semantic segmentation methods have an intrinsic limitation that training a model requires large amount of data with pixel-level annotations. To address this challenging issue, many researchers give attention to unsupervised domain adaptation for segmentation. Unsupervised seeks adapt the trained on source target domain. In paper, we introduce self-ensembling technique, one successful in classification. However, applying is very difficult because heavily-tuned manual...

Multi-resolution hash encoding has recently been proposed to reduce the computational cost of neural renderings, such as NeRF. This method requires accurate camera poses for renderings given scenes. However, contrary previous methods jointly optimizing and 3D scenes, naive gradient-based pose refinement using multi-resolution severely deteriorates performance. We propose a joint optimization algorithm calibrate learn geometric representation efficient encoding. Showing that oscillating...

Download This Paper Open PDF in Browser Add to My Library Share: Permalink Using these links will ensure access this page indefinitely Copy URL DOI

In this paper, we introduce a self-supervised approach for video object segmentation without human labeled data. Specifically, present Robust Pixel-level Matching Networks (RPM-Net), novel deep architecture that matches pixels between adjacent frames, using only color information from unlabeled videos training. Technically, RPM-Net can be separated in two main modules. The embedding module first projects input images into high dimensional space. Then the matching with deformable convolution...

While learning-based multi-view stereo (MVS) methods have recently shown successful performances in quality and efficiency, limited MVS data hampers generalization to unseen environments. A simple solution is generate various large-scale datasets, but generating dense ground truth for 3D structure requires a huge amount of time resources. On the other hand, if reliance on relaxed, systems will generalize more smoothly new To this end, we first introduce novel semi-supervised framework called...

Autonomous navigation in off-road conditions requires an accurate estimation of terrain traversability. However, traversability unstructured environments is subject to high uncertainty due the variability numerous factors that influence vehicle-terrain interaction. Consequently, it challenging obtain a generalizable model can accurately predict variety environments. This paper presents METAVerse, meta-learning framework for learning global and reliably predicts across diverse We train...

Recent unsupervised representation learning methods rely heavily on various transformations to generate distinctive views of given samples. Transformations for these are generally defined manually, requiring significant human effort design detailed configurations and validate practical efficacy. Furthermore, the diversity is quite limited in scope causing network be invariant only a small set data transformations. To address problems, we introduce neural transformation that learns diverse...