9 results about "Relative depth" patented technology

The invention discloses a transparent object depth completion method and system based on relative depth and mask attention, and belongs to the field of robots / artificial intelligence. The method comprises the following steps of: 1, acquiring multi-modal scene data and preprocessing the multi-modal scene data; 2, constructing a feature coding network based on mask attention and relative depth; 3, performing multi-feature fusion and full-image depth reconstruction on the feature coding network constructed in the step 2; and step 4, performing network training based on global loss optimization based on the multi-modal scene data preprocessed in the step 1 and reconstruction of the feature coding network in the step 3. The objective of the invention is to overcome the problem that the existing transparent object depth completion method is poor in accuracy and generalization in a real scene.

The invention relates to a gold thread height detection method, system and device for an optical module and a medium, and the method comprises the steps: carrying out the multi-scale convolution feature extraction of an image of the optical module, and fusing the V-channel features of the image into the convolution features; inputting the convolution feature map into a corresponding multi-scale detection head for gold thread positioning detection to obtain two-dimensional coordinate information of each gold thread; on the basis of the two-dimensional coordinate information, intercepting an interested area corresponding to each gold thread in the optical module image, and performing pixel-level depth detection on the interested area to obtain a relative depth map and a reference depth value of each gold thread; and performing abnormal value elimination on the relative depth map according to the Pauta criterion, and based on the obtained optimal relative depth map, combining the two-dimensional coordinate information and the corresponding reference depth value to obtain the actual height of the gold thread relative to the substrate. According to the invention, the gold thread height detection precision is improved, and especially the gold threads which are densely arranged, staggered and overlapped and are thin and long in shape have a relatively high detection effect.

The invention discloses a deblocking method based on explicit three-dimensional Gaussian scene layering, and the method comprises the steps: carrying out the constrained full-scene reconstruction through depth prior provided by a depth estimation base model, so as to obtain a three-dimensional Gaussian scene which meets the geometric requirements and is divided by a blocking object and a background; estimating the separation depth by utilizing the depth inconsistency between the two visual angles caused by the shielding object; learning and optimizing the separation depth of each view angle by using an adaptive weight learning method; and obtaining a non-shielding result by using the Gaussian of the separated depth splashing background. According to the invention, the depth estimation base model is used to provide the relative depth map with stronger geometric priori, and after the relative depth map is aligned with the actual scene depth, a scene with better geometry can be obtained through optimization; in the separation depth estimation part, the invention provides a method for determining the position and the separation depth of a shielding object by utilizing the inconsistency of the depths of different visual angles, and the optimization speed is greatly improved while various types of shielding are effectively removed.

This application discloses a 3D reconstruction method, storage medium, and computer program product, relating to the field of computer vision technology. The method includes: acquiring a sequence of unlabeled multi-frame video images; calling a feedforward 3D reconstruction model; obtaining ordinal depth loss, local affine invariant loss, and multi-view consistency constraint loss; the ordinal depth loss constrains the relative depth ordering of pixel pairs, the local affine invariant loss preserves locally aligned surface details, and the multi-view consistency constraint loss is used to determine the camera position and pose; jointly updating the model parameters using the ordinal depth loss, local affine invariant loss, and multi-view consistency constraint loss to obtain a 3D reconstruction optimization model; and inputting the multi-frame video image sequence into the 3D reconstruction optimization model for forward propagation to obtain 3D reconstruction information. By jointly updating the model parameters using three different loss functions, the reconstruction accuracy of dynamic scenes is improved without relying on expensive annotations and while preserving inference efficiency.

The application discloses a three-dimensional hand posture estimation method based on monocular RGB images, designs a feature enhancement method which can explicitly introduce the inherent skeleton structure of the hand, and adaptively enhances the features of the joint nodes to be estimated by using the associated information, so that the accuracy of hand posture estimation is finally improved. The designed method is as follows: firstly, convolutional neural network is used to extract joint node level semantic features and skeleton level semantic features from the input hand image, and a feature fusion module is used to cross semantic aggregation of the two features; then, the feature adaptive enhancement module can adaptively enhance the related features of each joint node by using the associated information thereof; then, the joint node two-dimensional heat map and the relative depth map are obtained through the output layer, and a multi-stage optimization strategy is adopted to continuously refine the two-dimensional heat map and the depth map to estimate more accurate hand two-dimensional joint node coordinates and relative depth; finally, the final hand three-dimensional coordinate information is calculated by using the camera parameters.

The application relates to a monocular camera absolute depth acquisition method, device, equipment and storage medium. The method comprises the following steps: acquiring a training image containing two ground features photographed by a monocular camera and the absolute distance between the two ground features in the training image; identifying the training image to acquire the pixel coordinates and the relative depth of the two ground features in the training image; obtaining the relative distance between the two ground features in the training image according to the pixel coordinates, the relative depth and the camera internal parameter matrix of the monocular camera; and obtaining the absolute depth of the two ground features in the training image according to the ratio of the absolute distance to the relative distance between the two ground features in the training image and the relative depth. The scheme provided by the application can reduce the influence of scene transformation on the absolute depth acquisition process and obtain accurate absolute depth information.

The invention discloses an interaction method, device and equipment based on intention perception and adaptive depth partitioning, and a computer readable storage medium. The interaction method comprises the following steps: acquiring a user shoulder key point coordinate to determine an original point and a reference scale of an interaction coordinate system; determining an entry threshold value and an exit threshold value of the Z axis based on the reference scale, wherein the entry threshold value is greater than the exit threshold value; obtaining the key point coordinates of the wrist of the user to calculate the Z-axis relative depth, the Z-axis speed and the X-Y plane speed of the wrist; switching a current interaction mode between a first interaction mode and a second interaction mode based on the depth, the speed and the threshold value; and according to the current interaction mode, generating a two-dimensional movement interaction instruction based on the X-Y plane displacement of the wrist, or generating a one-dimensional continuous parameter interaction instruction based on the Z-axis relative depth. The method has the advantages that mode flicker and mistaken touch are avoided, the interaction dimension is expanded, and the interaction stability and richness are improved.

Embodiments of the present application provide a depth estimation method and device, electronic equipment and storage medium, which are applied to electronic equipment, the electronic equipment comprising a camera device for shooting, the method comprising: obtaining a first image; inputting the first image into a depth estimation model to obtain a first depth image; obtaining a depth scale factor, wherein the depth scale factor is used to indicate a relationship between a relative depth of a pixel point on the first depth image and a depth value of the pixel point; and calculating depth information of the first depth image according to the depth scale factor and the first depth image. The actual depth value can be obtained without knowing the distance between left and right cameras of a binocular stereo camera in advance, and the actual depth value can be accurately known when using a monocular camera.