As the foundation of driverless vehicle and intelligent robots, Simultaneous Localization Mapping(SLAM) has attracted much attention these days. However, non-geometric modules traditional SLAM algorithms are limited by data association tasks have become a bottleneck preventing development SLAM. To deal with such problems, many researchers seek to Deep Learning for help. But most studies virtual datasets or specific environments, even sacrifice efficiency accuracy. Thus, they not practical...

Modeling scene geometry using implicit neural representation has revealed its advantages in accuracy, flexibility, and low memory usage. Previous approaches have demonstrated impressive results color or depth images but still difficulty handling poor light conditions large-scale scenes. Methods taking global point cloud as input require accurate registration ground truth coordinate labels, which limits their application scenarios. In this paper, we propose a new method that uses sparse LiDAR...

Directly learning multiple 3D objects motion from sequential images is difficult, while the geometric bundle adjustment lacks ability to localize invisible object centroid. To benefit both powerful understanding skill deep neural network meanwhile tackle precise geometry modeling for consistent trajectory estimation, we propose a joint spatial-temporal optimization-based stereo tracking method. From network, detect corresponding 2D bounding boxes on adjacent and regress an initial box. Dense...

For a long time, the point cloud completion task has been regarded as pure generation task. After obtaining global shape code through encoder, complete is generated using priorly learnt by networks. However, such models are undesirably biased towards prior average objects and inherently limited to fit geometry details. In this letter, we propose Graph-Guided Deformation Network, which respectively regards input data intermediate controlling supporting points, optimization guided graph...

Point clouds collected by real-world sensors are always unaligned and sparse, which makes it hard to reconstruct the complete shape of object from a single frame data. In this work, we manage provide point sparse input with pose disturbance limited translation rotation. We also use temporal information enhance completion model, refining output sequence inputs. With help gated recovery units(GRU) attention mechanisms as units, propose cloud framework that accepts inputs, outputs consistent...

Mesh plays an indispensable role in dense realtime reconstruction essential robotics. Efforts have been made to maintain flexible data structures for 3D fusion, yet efficient incremental framework specifically designed online mesh storage and manipulation is missing. We propose a novel compactly generate, update, refine scene upon volumetric representation. Maintaining spatial-hashed field of cubes, we distribute vertices with continuous value on discrete edges that support <i...

Semantic mapping based on the supervised object detectors is sensitive to image distribution. In real-world environments, detection and segmentation performance can lead a major drop, preventing use of semantic in wider domain. On other hand, development vision-language foundation models demonstrates strong zero-shot transferability across data It provides an opportunity construct generalizable instance-aware maps. Hence, this work explores how boost from generated models. We propose...

Robust visual localization for urban vehicles remains challenging and unsolved. The limitation of computation efficiency memory size has made it harder large-scale applications. Since semantic information serves as a stable compact representation the environment, we propose coarse-to-fine system based on map. Pole-like objects are stored in map, then extracted from semantically segmented images observations. Localization is performed by particle filter, followed pose alignment module...

Mesh plays an indispensable role in dense real-time reconstruction essential robotics. Efforts have been made to maintain flexible data structures for 3D fusion, yet efficient incremental framework specifically designed online mesh storage and manipulation is missing. We propose a novel compactly generate, update, refine scene upon volumetric representation. Maintaining spatial-hashed field of cubes, we distribute vertices with continuous value on discrete edges that support O(1) vertex...

In the practical application of point cloud completion tasks, real data quality is usually much worse than CAD datasets used for training. A small amount noisy will significantly impact overall system's accuracy. this paper, we propose a evaluation network to score clouds and help judge before applying model. We believe our scoring method can researchers select more appropriate subsequent reconstruction avoid manual parameter adjustment. Moreover, model fast straightforward be directly...

Semantic mapping based on the supervised object detectors is sensitive to image distribution. In real-world environments, detection and segmentation performance can lead a major drop, preventing use of semantic in wider domain. On other hand, development vision-language foundation models demonstrates strong zero-shot transferability across data It provides an opportunity construct generalizable instance-aware maps. Hence, this work explores how boost from generated models. We propose...

Dynamic Gaussian splatting has led to impressive scene reconstruction and image synthesis advances in novel views. Existing methods, however, heavily rely on pre-computed poses initialization by Structure from Motion (SfM) algorithms or expensive sensors. For the first time, this paper addresses issue integrating self-supervised VO into our pose-free dynamic method (VDG) boost pose depth static-dynamic decomposition. Moreover, VDG can work with only RGB input construct scenes at a faster...

Robust visual localization for urban vehicles remains challenging and unsolved. The limitation of computation efficiency memory size has made it harder large-scale applications. Since semantic information serves as a stable compact representation the environment, we propose coarse-to-fine system based on map. Pole-like objects are stored in map, then extracted from semantically segmented images observations. Localization is performed by particle filter, followed pose alignment module...

The image-based 3D object detection task expects that the predicted bounding box has a “tightness” projection (also referred to as cuboid) facilitate 2D-based training, which fits contour well on image while remaining reasonable space. These requirements bring significant challenges annotation. Projecting Lidar-labeled boxes leads non-trivial misalignment, directly drawing cuboid cannot access original information. In this work, we propose learning-based adaptation approach automatically...

Point clouds collected by real-world sensors are always unaligned and sparse, which makes it hard to reconstruct the complete shape of object from a single frame data. In this work, we manage provide point sparse input with pose disturbance limited translation rotation. We also use temporal information enhance completion model, refining output sequence inputs. With help gated recovery units(GRU) attention mechanisms as units, propose cloud framework that accepts inputs, outputs consistent...

Modeling scene geometry using implicit neural representation has revealed its advantages in accuracy, flexibility, and low memory usage. Previous approaches have demonstrated impressive results color or depth images but still difficulty handling poor light conditions large-scale scenes. Methods taking global point cloud as input require accurate registration ground truth coordinate labels, which limits their application scenarios. In this paper, we propose a new method that uses sparse LiDAR...

In the practical application of point cloud completion tasks, real data quality is usually much worse than CAD datasets used for training. A small amount noisy will significantly impact overall system's accuracy. this paper, we propose a evaluation network to score clouds and help judge before applying model. We believe our scoring method can researchers select more appropriate subsequent reconstruction avoid manual parameter adjustment. Moreover, model fast straightforward be directly...

IoT has long been a hot topic, and many researchers have made great efforts especially in wireless charging communication between smart furniture, recognization localization. However, people always divide charging, communicating sensing apart. In our paper, we put forward simple but practical method to make use of NFC resonance-type WPT build up system which can be used locate, identify, charge devices while providing platform for objects communicate with each other. Our acts as an...

Directly learning multiple 3D objects motion from sequential images is difficult, while the geometric bundle adjustment lacks ability to localize invisible object centroid. To benefit both powerful understanding skill deep neural network meanwhile tackle precise geometry modeling for consistent trajectory estimation, we propose a joint spatial-temporal optimization-based stereo tracking method. From network, detect corresponding 2D bounding boxes on adjacent and regress an initial box. Dense...

A robust 3D object tracker which continuously tracks surrounding objects and estimates their trajectories is key for self-driving vehicles. Most existing tracking methods employ a tracking-by-detection strategy, usually requires complex pair-wise similarity computation neglects the nature of continuous motion. In this paper, we propose to directly learn correspondences from temporal point cloud data infer motion information correspondence patterns. We modify standard detector process two...

The image-based 3D object detection task expects that the predicted bounding box has a ``tightness'' projection (also referred to as cuboid), which fits contour well on image while still keeping geometric attribute space, e.g., physical dimension, pairwise orthogonal, etc. These requirements bring significant challenges annotation. Simply projecting Lidar-labeled boxes leads non-trivial misalignment, directly drawing cuboid cannot access original information. In this work, we propose...

For a long time, the point cloud completion task has been regarded as pure generation task. After obtaining global shape code through encoder, complete is generated using priorly learnt by networks. However, such models are undesirably biased towards prior average objects and inherently limited to fit geometry details. In this paper, we propose Graph-Guided Deformation Network, which respectively regards input data intermediate controlling supporting points, optimization guided graph...