[time]
2022-07-08T11:22:18+02:00

[track]
2

[team_name]
SZUSCRI

[team_institution]
Shenzhen University Smart City Research Institute

[logolink]


[team_members]
Yusong Li, Jiawei Wan, Shengjun Tang

[reference_person]
Yusong Li, Jiawei Wan, Shengjun Tang

[reference_email]
IPIN2022@wjwent.onmicrosoft.com

[description_short]
System specification 
Data Preprocessing 
After receiving image input: 
1. A search is performed first, matching the query image to a database image to perform a rough map-level search. 
2. Then perform common clustering on the search results. 
3. Then perform local feature matching. For each location, match the 2D key points detected in the query image with the 3D points contained in the location in turn. 
Positioning Phase 
1. First, the CNN model will be used to predict the dense features of the query image and the reference image in the matching sequence. 
2. Obtain local 3D points and initial camera poses according to the features of the previous step. 
3. Calculate the error between the query image and the reference image. 
4. Use the optimizer to align the features to adjust the camera pose. 
5. Continuously iterative optimization with gradient-based optimization algorithm. 
6. Finally, output the camera pose with the smallest error. 


[description_long_link]
https://wjwent-my.sharepoint.com/:b:/g/personal/ipin2022_wjwent_onmicrosoft_com/EVcV_H9b25FLsXhCChl_uZgBmUo84kZbAcQoBu1_peVxug?e=B7IzO0

[publish_check_]
true

[results_check_]
true

[data_check_]
true

[pdf_check_]
false

[video_check_]
false