Convolutional neural network-based approaches have achieved remarkable progress in semantic segmentation. However, these heavily rely on annotated data which are labor intensive. To cope with this limitation, automatically generated from graphic engines used to train segmentation models. the models trained synthetic difficult transfer real images. tackle issue, previous works considered directly adapting source unlabeled target (to reduce inter-domain gap). Nonetheless, techniques do not...

We propose a novel video object segmentation algorithm based on pixel-level matching using Convolutional Neural Networks (CNN). Our network aims to distinguish the target area from background basis of similarity between two units. The proposed represents features different depth layers in order take advantage both spatial details and category-level semantic information. Furthermore, we feature compression technique that drastically reduces memory requirements while maintaining capability...

ResNet or DenseNet? Nowadays, most deep learning based approaches are implemented with seminal backbone networks, among them the two arguably famous ones and DenseNet. Despite their competitive performance overwhelming popularity, inherent drawbacks exist for both of them. For ResNet, identity shortcut that stabilizes training might limit its representation capacity, DenseNet mitigates it multi-layer feature concatenation. However, dense concatenation causes a new problem requiring high GPU...

Calibration of wide field-of-view cameras is a fundamental step for numerous visual media production applications, such as 3D reconstruction, image undistortion, augmented reality and camera motion estimation. However, existing calibration methods require multiple images pattern (typically checkerboard), assume the presence lines, manual interaction and/or need an sequence. In contrast, we present novel fully automatic deep learning-based approach that overcomes all these limitations works...

One of the most hazardous driving scenario is overtaking a slower vehicle, indeed, in this case front vehicle (being overtaken) can occlude an important part field view rear vehicle's driver. This lack visibility probable cause accidents context. Recent research works tend to prove that augmented reality applied assisted significantly reduce risk accidents. In paper, we present real-time marker-less system see through cars. For purpose, two cars are equipped with cameras and appropriate...

To reconstruct a 3D scene from set of calibrated views, traditional multi-view stereo techniques rely on two distinct stages: local depth maps computation and global fusion. Recent studies concentrate deep neural architectures for estimation by using conventional fusion method or direct reconstruction network regressing Truncated Signed Distance Function (TSDF). In this paper, we advocate that replicating the stages framework with networks improves both interpretability accuracy results. As...

Segment anything model (SAM) developed by Meta AI Research has recently attracted significant attention. Trained on a large segmentation dataset of over 1 billion masks, SAM is capable segmenting any object certain image. In the original work, authors turned to zero-short transfer tasks (like edge detection) for evaluating performance SAM. Recently, numerous works have attempted investigate in various scenarios recognize and segment objects. Moreover, projects emerged show versatility as...

Generative AI (AIGC, a.k.a. generated content) has made remarkable progress in the past few years, among which text-guided content generation is most practical one since it enables interaction between human instruction and AIGC. Due to development text-to-image as well 3D modeling technologies (like NeRF), text-to-3D become a newly emerging yet highly active research field. Our work conducts first comprehensive survey on help readers interested this direction quickly catch up with its fast...

Rotating and zooming cameras, also called PTZ (Pan-Tilt-Zoom) are widely used in modern surveillance systems. While their ability allows acquiring detailed images of the scene, it makes calibration more challenging since any action results a modification intrinsic parameters. Therefore, such camera has to be computed online; this process is self-calibration. In paper, given an image pair captured by camera, we propose deep learning based approach automatically estimate focal length...

For scalable autonomous driving, a robust map-based localization system, independent of GPS, is fundamental. To achieve such localization, online high-definition (HD) map construction plays significant role in accurate estimation the pose. Although recent advancements HD have predominantly investigated on vectorized representation due to its effectiveness, they suffer from computational cost and fixed parametric model, which limit scalability. alleviate these limitations, we propose novel...

This paper presents a robust approach for road marking detection and recognition from images captured by an embedded camera mounted on car. Our method is designed to cope with illumination changes, shadows, harsh meteorological conditions. Furthermore, the algorithm can effectively group complex multi-symbol shapes into individual marking. For this purpose, proposed technique relies MSER features obtain candidate regions which are further merged using density-based clustering. Finally, these...

In this paper, we propose a noise-aware exposure control algorithm for robust robot vision. Our method aims to capture best-exposed images, which can boost the performance of various computer vision and robotics tasks. For purpose, carefully design an image quality metric that captures complementary attributes ensures light-weight computation. Specifically, our consists combination gradient, entropy, noise metrics. The synergy these measures allows preservation sharp edges rich texture in...

Convolutional neural network-based approaches have achieved remarkable progress in semantic segmentation. However, these heavily rely on annotated data which are labor intensive. To cope with this limitation, automatically generated from graphic engines used to train segmentation models. the models trained synthetic difficult transfer real images. tackle issue, previous works considered directly adapting source unlabeled target (to reduce inter-domain gap). Nonetheless, techniques do not...

Inferring traffic object such as lane information is of foremost importance for deployment autonomous driving. Previous approaches focus on offline construction HD map inferred with GPS localization, which insufficient globally scalable To alleviate these issues, we propose online learning framework that detects elements from onboard sensor observations. We represent the a graph; InstaGraM, instance-level graph modeling brings accurate and fast end-to-end vectorized learning. Along strategy,...

In most of computer vision applications, motion blur is regarded as an undesirable artifact. However, it has been shown that in image may have practical interests fundamental problems. this work, we propose a novel framework to estimate optical flow from single motion-blurred end-to-end manner. We design our network with transformer networks learn globally and locally varying motions encoded features input, decode left right frame without explicit supervision. A estimator then used the...

Estimating the motion of camera together with 3D structure scene from a monocular vision system is complex task that often relies on so-called rigidity assumption. When observing dynamic environment, this assumption violated which leads to an ambiguity between ego-motion and objects. To solve problem, we present self-supervised learning framework for object field estimation videos. Our contributions are two-fold. First, propose two-stage projection pipeline explicitly disentangle motions...

Drivable region detection is challenging since various types of road, occlusion or poor illumination condition have to be considered in a outdoor environment, particularly at night. In the past decade, Many efforts been made solve these problems, however, most already existing methods are designed for visible light cameras, which inherently inefficient under low conditions. this paper, we present drivable algorithm thermal-infrared cameras order overcome aforementioned problems. The novelty...

Abrupt motion of camera or objects in a scene result blurry video, and therefore recovering high quality video requires two types enhancements: visual enhancement temporal upsampling. A broad range research attempted to recover clean frames from blurred image sequences temporally upsample by interpolation, yet there are very limited studies handling both problems jointly. In this work, we present novel framework for deblurring, interpolating extrapolating sharp motion-blurred an end-to-end...

We present a new linear method for RGB-D based simultaneous localization and mapping (SLAM). Compared to existing techniques relying on the Manhattan world assumption defined by three orthogonal directions, our approach is designed more general scenario of Atlanta world. It consists vertical direction set horizontal directions thus can represent wider range scenes. Our leverages structural regularity decouple non-linearity camera pose estimations. This allows us separately estimate rotation...