Recently, transformers have been widely adopted for various computer vision tasks and show promising results due to their ability to encode long-range spatial dependencies in an image effectively. However, very few studies on adopting transformers in self-supervised depth estimation have been conducted. When replacing the CNN architecture with the transformer in self-supervised learning of depth, we encounter several problems such as problematic multi-scale photometric loss function when used with transformers and, insufficient ability to capture local details. In this letter, we propose an attention-based decoder module, Pixel-Wise Skip Attention (PWSA), to enhance fine details in feature maps while keeping global context from transformers. In addition, we propose utilizing self-distillation loss with single-scale photometric loss to alleviate the instability of transformer training by using correct training signals. We demonstrate that the proposed model performs accurate predictions on large objects and thin structures that require global context and local details. Our model achieves state-of-the-art performance among the self-supervised monocular depth estimation methods on KITTI and DDAD benchmarks.

Pedestrian detection is a critical task for safety-critical systems, but detecting pedestrians is challenging in low-light and adverse weather conditions. Thermal images can be used to improve robustness by providing complementary information to RGB images. Previous studies have shown that multi-modal feature fusion using convolution operation can be effective, but such methods rely solely on local feature correlations, which can degrade the performance capabilities. To address this issue, we propose an attention-based novel fusion network, referred to as INSANet (INtra-INter Spectral Attention Network), that captures global intra- and inter-information. It consists of intra- and inter-spectral attention blocks that allow the model to learn mutual spectral relationships. Additionally, we identified an imbalance in the multispectral dataset caused by several factors and designed an augmentation strategy that mitigates concentrated distributions and enables the model to learn the diverse locations of pedestrians. Extensive experiments demonstrate the effectiveness of the proposed methods, which achieve state-of-the-art performance on the KAIST dataset and LLVIP dataset. Finally, we conduct a regional performance evaluation to demonstrate the effectiveness of our proposed network in various regions.

In this paper, multispectral pedestrian detection is mainly discussed, which can contribute to assigning human-aware properties to automated forklifts to prevent accidents, such as collisions, at an early stage. Since there was no multispectral pedestrian detection dataset in an intralogistics domain, we collected a dataset; the dataset employs a method that aligns image pairs with different domains, i.e. RGB and thermal, without the use of a cumbersome device such as a beam splitter, but rather by exploiting the disparity between RGB sensors and camera geometry. In addition, we propose a multispectral pedestrian detector called SSD 2.5D that can not only detect pedestrians but also estimate the distance between an automated forklift and workers. In extensive experiments, the performance of detection and centroid localization is validated with respect to evaluation metrics used in the driving car domain but with distinct categories, such as hazardous zone and warning zone, to make it more applicable to the intralogistics domain.

Multispectral image alignment plays a crucial role in exploiting complementary information between different spectral images. Homography-based image alignment can be a practical solution considering a tradeoff between runtime and accuracy. Existing methods, however, have difficulty with multispectral images due to the additional spectral gap or require expensive human labels to train models. To solve these problems, this paper presents a comprehensive study on multispectral homography estimation in an unsupervised learning manner. We propose a curriculum data augmentation, an effective solution for models learning spectrum-agnostic representation by providing diverse input pairs. We also propose to use the phase congruency loss that explicitly calculates the reconstruction between images based on low-level structural information in the frequency domain. To encourage multispectral alignment research, we introduce a novel FLIR corresponding dataset that has manually labeled local correspondences between multispectral images. Our model achieves state-of-the-art alignment performance on the proposed FLIR correspondence dataset among supervised and unsupervised methods while running at 151 FPS . Furthermore, our model shows good generalization ability on the M3FD dataset without finetuning.

Scene text detection in industrial environments is challenging due to low-contrast, corrosion, and glare on metallic surfaces, which affect detection accuracy. Furthermore, symbols such as engraved logos resemble text, leading to misclassification and increased false positive rates. In this study, we propose a framework that applies synthetic data augmentation to handle low-contrast conditions and employs contrastive learning to enhance text-symbol differentiation. This framework improves robustness and generalization by training the model to diverse low-contrast scenarios. Furthermore, it reduces misclassification by using false positives as negative samples, enhancing the separation between text and symbols. Since the proposed method is adopted only during the training phase, it avoids computational overhead during inference. Moreover, it does not rely on modifications to the architecture, allowing flexible integration into existing text detectors. In addition, we observed ambiguities in the annotation criteria of the MPSC dataset, particularly in distinguishing individual text instances and labeling their ground-truth bounding boxes. To resolve this, we re-annotate the dataset by defining a consistent standard for text instance labeling. Our experimental results demonstrate that the proposed model achieves state-of-the-art performance on the MPSC dataset, validating its applicability to industrial text detection tasks.

For autonomous systems operating in degraded visual environments, robust 3D perception depends on accurate, metric-scale depth prediction. Yet, thermal depth completion—despite being well-suited for such conditions—remains relatively underexplored. Conventional depth completion approaches, which fuse sparse LiDAR with RGB images, perform well under well-lit conditions but break down in low light or adverse conditions due to the inherent limitations of RGB sensing. We present the first diffusion-based zero-shot thermal depth completion framework, extending the capability of an RGB-trained foundation model to the thermal domain. Our approach is twofold: First, to bridge the domain gap, we introduce a Thermal-Aware Multispectral Adaptation (TAMA) that leverages paired RGB-thermal images to enhance the model’s understanding of thermal image. Second, we replace the iterative optimization process with our Thermal-aware Proxy Alignment (TAPA) module, which directly integrates thermal guidance and sparse metric constraints into the diffusion sampling process. Our framework enables reliable, metric-scale depth prediction across diverse thermal scenes. Notably, our zero-shot method demonstrates performance competitive with fully-supervised baselines that require task-specific training, establishing a foundation for robust 3D perception in challenging environments.