SyreaNet: A Physically Guided Underwater Image Enhancement Framework Integrating Synthetic and Real Images
J. Wen, Benwen Chen
CodeRestorationData SynthesisDomain AdaptationImagePhysics Related
🎯 What it does: Proposed the SyreaNet framework, which integrates a synthesis module based on a revised model with a physics-guided decoupled network for seawater image enhancement;
🎯 What it does: Proposes a test-time domain adaptation framework for monocular depth estimation that can instantly adapt the source pre-trained model to test data in a source-free and unsupervised manner.
The Reflectance Field Map: Mapping Glass and Specular Surfaces in Dynamic Environments
P. Foster, B. Kuipers
CodeComputational EfficiencySimultaneous Localization and MappingPoint Cloud
🎯 What it does: Proposed a LiDAR-based reflectance field map that can real-time detect mirror-reflective surfaces such as glass, metal, and mirrors, integrating light field mapping theory with occupancy grid mapping.
TODE-Trans: Transparent Object Depth Estimation with Transformer
Kan Chen, Bin Li
CodeDepth EstimationTransformerImage
🎯 What it does: Developed a Transformer-based method for depth estimation of transparent objects, using a single RGB-D input to predict the surface depth of transparent objects
🎯 What it does: Proposed and verified a robust sim-to-real transmission pipeline for collecting full-body shape information of soft robots under high-fidelity point cloud representations, and evaluated the model directly on real internal camera images after training on simulated data.
Transferring Implicit Knowledge of Non-Visual Object Properties Across Heterogeneous Robot Morphologies
Gyan Tatiya, Jivko Sinapov
CodeClassificationRecognitionRobotic Intelligence
🎯 What it does: Propose a multi-stage projection framework that can transfer implicit object attribute knowledge between different robot morphologies, evaluate it on object attribute recognition and identity recognition tasks, and introduce data augmentation techniques to enhance model generalization capabilities.
🎯 What it does: Proposed a fast monocular depth estimation method called UDepth, aiming to provide 3D perception capabilities for low-cost underwater robots.
🎯 What it does: Propose the EvLPSNet network to address the uncertainty-aware panoptic segmentation problem for LiDAR point clouds, predicting semantic, instance segmentation, and uncertainty estimation for each point.
🎯 What it does: Proposes an unsupervised RGB-to-thermal domain adaptation method that utilizes a multi-domain attention network to achieve thermal image classification and semantic segmentation.
🎯 What it does: Proposed an open-source adversarial scenario generator named V2XP-ASG for generating realistic and challenging traffic scenarios for LiDAR-based multi-agent perception systems.
Viewer-Centred Surface Completion for Unsupervised Domain Adaptation in 3D Object Detection
Darren Tsai, Stewart Worrall
CodeObject DetectionDomain AdaptationPoint Cloud
🎯 What it does: Propose a view-based surface completion network (VCN) that unifies 3D detection object representations across different datasets under an unsupervised domain adaptation framework, to reduce performance degradation caused by differences in LiDAR scanning modes.
🎯 What it does: Proposed a lightweight Vision-and-Pointcloud Transformer (ViPFormer) for unsupervised point cloud understanding, and pre-trained it to migrate to downstream tasks such as 3D shape classification and semantic segmentation.
🎯 What it does: A visual pitch and roll angle estimation method for inland waterway vessels was developed, utilizing CNN for water surface segmentation, stereo vision reconstruction, and geometric calculations to estimate pitch and roll.
Vitreoretinal Surgical Robotic System with Autonomous Orbital Manipulation using Vector-Field Inequalities
Yuki Koyama, K. Harada
CodeRobotic Intelligence
🎯 What it does: Proposes a method for autonomous ocular orbit manipulation in robot-assisted vitreoretinal surgery, utilizing a remote control system to achieve eye rotation for expanding the retinal field of view
🎯 What it does: Proposes a target-guided knowledge distillation framework for weakly supervised referring expression localization, leveraging target-related prediction information from a pre-trained teacher model to guide student model training, thereby enhancing weakly supervised localization performance.
🎯 What it does: Proposes a zero-shot object detection method based on dynamic semantic vectors, and designs a bidirectional classification branch network with an optimization process incorporating the N-pair loss