Some discussions and questions about the tracking accuracy of NICE slam #126
Description
Hello, this is a great work presented at CVPR. I have learned a lot from the detailed designs in your excellent work. However, when I try to run NICE-SLAM or implement a neural implicit SLAM, I have encountered many problems, especially during the tracking process.
The first frame can be easily remembered by the network, but I often find that even a slight camera rotation causes the tracking error to increase significantly.
I tried using the configuration provided in the repository to run the SLAM on Replica data, but the results are still poor, even in simple scenes such as Room0 and Room1.
The detials are showed below.
The results are not as good as the output/vis/Replica/room1 showed.
maybe I have miss some important settings? I am so confused, please help me.