WiMi Hologram Cloud Inc (WIMI)

[StockpickerCAPC]

WIMI Develops A 3D Object Recognition System Based On Multi-View Feature Fusion.

Source
https://www.newstrail.com/wimi-develops-a-3dobject-recognition-system-based-on-multi-view-feature-fusion/

February 153, 2023

3D object recognition, as the basis of 3D data understanding and analysis, has become an important research direction in the field of computer vision. With the development of computer technology, the application of 3D object recognition technology is more and more extensive. For example, robots use 3D object recognition technology to model and understand the 3D scene, face recognition system processes images and depth information through 3D object recognition, augmented reality, and virtual reality use 3D object recognition technology to generate and render 3D objects, and so on.

It is said that WIMI Hologram Cloud INC. (NASDAQ: WIMI) has developed a 3D object recognition system based on multi-view feature fusion. It uses a convolutional neural network to analyze different perspectives of 3D objects and infuses features of multiple views to infer the global information of 3D objects, which is then input into the fully connected network for classification. Extrapolating labels for three-dimensional objects from multiple views.

WIMI’s 3D object recognition system based on multi-view feature fusion mainly consists of three parts:
- view information selection,
- feature extraction, and
- feature fusion.
The perspective information module is mainly used to project three-dimensional objects into two-dimensional planes from multiple angles. Different perspectives involve different directions and structure information of objects. Graph structure can be established among multiple views, and clustering can be done according to spatial distribution. RA’s reasonable view information selection strategy can optimize network training data.

Feature extraction module mainly uses a convolutional neural network to extract features. The feature mapping module can be applied to the feature response graph of the view behind the convolution layer, and the ma multi-layer perception machine is used to learn multiple mapping matrices, which map the corresponding view to an approximate feature space respectively. The mapping matrix can summarize the perspective transformation relationship between the views, and map the feature map to a group-level feature describing the region.

Feature fusion module mainly uses reasonable and effective strategies to fuse multiple features and realizes multi-layer fusion based on clustering. The convolution operation is used to wweighthe features of high-dimensional views and encode the weight information between different views. Convolution processes feature response graphs with spatial information. Features are extracted from the convolutional layer of the convolutional neural network and maximum pooling is used to obtain the maximum response on the feature graph. Moreover, cthe orrelation between adjacent views is learned to generate global features with mthe ore descriptive ability and fused into feature graphs.

After all view features are fused into global features, the global features are input into the full connection layer, ahe high-dimensional features among the fusion features with spatial information are mmine and the classification and output results are completed.

Three-dimensional object recognition technology is one of the core technologies of computer vision, which is the key technology of three-dimensional scene understanding and the basis for the machines to understand and interact with the world. It has extremely broad application prospects in automatic driving, intelligent robots, intelligent transportation, autonomous navigation, and other fields. WIMI will also continue to expand the application field of its 3D object recognition algorithm based on multi-view feature fusion.