Caoang327
Incorrect Timings in the Paper? Table 8
The paper introduction mentions that the speed of the FWD model is better than existing methods. However, Table 8 reports order(s) of magnitude slowdown vs PixelNerf and IBRNet.
Is the table accurate? This means FWD would take upwards of 30 seconds to generate a novel view from 3 views. Another strange thing about this table is that the runtimes decrease as number of views increases from 3 to 6.
Hi. Thanks for the pointing. This table is not totally correct: the "time" here is not real-time, it is actually FPS. That's why larger views will lead to smaller numbers/FPS. I originally designed this table using "time" and later changed it to "FPS" so that it would be consistent to the main table. In the last column, I forgot to delete the "min" in "0.009min"
