Closed ruoshiliu closed 2 years ago
Hi @ruoshiliu , yeah we run the above-mentioned on CPU - developing a custom CUDA kernel for this will for sure speed things up quite a bit! If you choose an appropriate depth_range, you can also drop the use_cube_intersection function; we used it such that no tuning of range is required. Best of luck for your research !
That makes sense. Thank you for your suggestion!
Hi, I was running your code on a custom dataset and I noticed the GPU utilization rate is pretty low due to the method
check_ray_intersection_with_unit_cube
bottlenecking the CPU. First of all, do you notice this during your experiments? What's a typical GPU utilization rate for you? If this is the case for you as well, I was wondering if settinguse_cube_intersection: false
and usingdepth_range
will potentially hurt the model performance?Would really appreciate if you can share some insights on this. Thanks in advanced!