Question about FPS in inference about scenerf

[anhquancao]

Hello @haiphamcse, I haven't calculated the frames per second (FPS) yet, but the inference time is primarily influenced by the number of views you'd like to generate. Each view typically requires about 2 seconds for generation. Additionally, the tsdf fusion process also scales linearly with the number of views.

However, the inference time can be significantly reduced by directly extracting the occupancy from the density volume, similar to the approach demonstrated in BTS paper: https://fwmb.github.io/bts/

[haiphamcse]

Hi there, thank you for your reply. Have the team done any test on the way to extract the occupancy out other than the TSDF fusion of novel views yet? To my best knowledge, to obtain the results in the paper 13.84 IoU we need to generate a lot of novel depth (~27 frames), it is necessary or a smaller amount of frames will still give comparable results

[anhquancao]

Hello @haiphamcse, I haven't experimented with obtaining the occupancy directly from the density field. In the case of SemKITTI, rotating views (angle != 0) only yielded marginal improvements, likely because the dataset mostly consists of front-facing images. I remembered sampling only 10 views with an angle of 0 and an interval of 1 meter also give roughly the same performance."

[haiphamcse]

Thank you for your reply. I will close the issue now.