We present a novel method for local image feature matching. Instead of performing detection, description, and matching sequentially, we propose to first establish pixel-wise dense matches at coarse level later refine the good fine level. In contrast methods that use cost volume search correspondences, self cross attention layers in Transformer obtain descriptors are conditioned on both images. The global receptive field provided by enables our produce low-texture areas, where detectors...

We present a novel framework named NeuralRecon for real-time 3D scene reconstruction from monocular video. Unlike previous methods that estimate single-view depth maps separately on each key-frame and fuse them later, we propose to directly reconstruct local surfaces represented as sparse TSDF volumes video fragment sequentially by neural network. A learning-based fusion module based gated recurrent units is used guide the network features fragments. This de-sign allows capture smoothness...

Recent advances in implicit neural representations and differentiable rendering make it possible to simultaneously recover the geometry materials of an object from multi-view RGB images captured under unknown static illumination. Despite promising results achieved, indirect illumination is rarely modeled previous methods, as requires expensive recursive path tracing which makes inverse computationally intractable. In this paper, we propose a novel approach efficiently recovering...

We propose a new method named OnePose for object pose estimation. Unlike existing instance-level or category-level methods, does not rely on CAD models and can handle objects in arbitrary categories without instance-or category-specific network training. draws the idea from visual localization only requires simple RGB video scan of to build sparse SfM model object. Then, this is registered query images with generic feature matching network. To mitigate slow runtime we graph attention that...

We are witnessing an explosion of neural implicit representations in computer vision and graphics. Their applicability has recently expanded beyond tasks such as shape generation image-based rendering to the fundamental problem 3D reconstruction. However, existing methods typically assume constrained environments with constant illumination captured by a small set roughly uniformly distributed cameras. introduce new method that enables efficient accurate surface reconstruction from Internet...

Underwater image enhancement presents a significant challenge due to the complex and diverse underwater environments that result in severe degradation phenomena such as light absorption, scattering, color distortion. More importantly, obtaining paired training data for these scenarios is challenging task, which further hinders generalization performance of models. To address issues, we propose novel approach, Hybrid Contrastive Learning Regularization (HCLR-Net). Our method built upon...

In this paper, we propose a novel system named Disp R-CNN for 3D object detection from stereo images. Many recent works solve problem by first recovering point cloud with disparity estimation and then apply detector. The map is computed the entire image, which costly fails to leverage category-specific prior. contrast, design an instance network (iDispNet) that predicts only pixels on objects of interest learns shape prior more accurate estimation. To address challenge scarcity annotation in...

We propose a new method for object pose estimation without CAD models. The previous feature-matching-based OnePose has shown promising results under one-shot setting which eliminates the need models or object-specific training. However, relies on detecting repeatable image keypoints and is thus prone to failure low-textured objects. keypoint-free pipeline remove keypoint detection. Built upon detector-free feature matching LoFTR, we devise SfM reconstruct semi-dense point-cloud model object....

We present a novel framework named NeuralRecon for real-time 3D scene reconstruction from monocular video. Unlike previous methods that estimate single-view depth maps separately on each key-frame and fuse them later, we propose to directly reconstruct local surfaces represented as sparse TSDF volumes video fragment sequentially by neural network. A learning-based fusion module based gated recurrent units is used guide the network features fragments. This design allows capture smoothness...

In this paper, we propose a novel system named Disp R-CNN for 3D object detection from stereo images. Many recent works solve problem by first recovering point clouds with disparity estimation and then apply detector. The map is computed the entire image, which costly fails to leverage category-specific prior. contrast, design an instance network (iDispNet) that predicts only pixels on objects of interest learns shape prior more accurate estimation. To address challenge scarcity annotation...

We present a novel method for local image feature matching. Instead of performing detection, description, and matching sequentially, we propose to first establish pixel-wise dense matches at coarse level later refine the good fine level. In contrast methods that use cost volume search correspondences, self cross attention layers in Transformer obtain descriptors are conditioned on both images. The global receptive field provided by enables our produce low-texture areas, where detectors...

This paper presents an approach that reconstructs a hand-held object from monocular video. In contrast to many recent methods directly predict geometry by trained network, the proposed does not require any learned prior about and is able recover more accurate detailed geometry. The key idea hand motion naturally provides multiple views of can be reliably estimated pose tracker. Then, recovered solving multi-view reconstruction problem. We devise implicit neural representation-based method...

We present a novel method for local image feature matching. Instead of performing detection, description, and matching sequentially, we propose to first establish pixel-wise dense matches at coarse level later refine the good fine level. In contrast methods that use cost volume search correspondences, self cross attention layers in Transformer obtain descriptors are conditioned on both images. The global receptive field provided by enables our produce low-texture areas, where detectors...

This paper targets high-fidelity and real-time view synthesis of dynamic 3D scenes at 4K resolution. Recently, some methods on have shown impressive rendering quality. However, their speed is still limited when high-resolution images. To overcome this problem, we propose 4K4D, a 4D point cloud representation that supports hardware rasterization enables unprecedented speed. Our built feature grid so the points are naturally regularized can be robustly optimized. In addition, design novel...