Chris Offner (@chrisoffner3d)
2024-10-06 | โค๏ธ 130 | ๐ 13
Here is the output of DepthCrafter (Hu, Gao, Li et al., 2024). Very smooth! โญ :)
However, this model conditions on a full video, i.e. all frames need to be known ahead of time โ making it unsuitable for online use in robotics. Itโs also very computationally expensive.
Unfortunately the code base is hardcoded for CUDA, making it impossible for me to run it on MPS on my M3 Max with 128 GB shared memory without major code modifications.
On the RTX 3080 Ti with 12GB memory I have access to, it ran out of memory in its full-resolution mode. Running it at reduced resolution of 512 worked but took quite a while as well.
๋ฏธ๋์ด
Quoted: @chrisoffner3d
Again, when thereโs a lot of sky in the frame, all bets are off. The max depth oscillates between 10,000 (infinity) and values down to <200 in successive frames. https://x.com/chrisoffner3d/status/1842516573078135154/video/1โฆ
์ธ์ฉ ํธ์
Chris Offner (@chrisoffner3d)
Again, when thereโs a lot of sky in the frame, all bets are off. The max depth oscillates between 10,000 (infinity) and values down to <200 in successive frames. https://t.co/rkD7oAB4zt
๐ฌ ์์
