FPS Reported In Paper

[ghost]

Hello,

Many thanks for this excellent work! I was very impressed by your first Pseudo-LiDAR paper and now your team's idea to include a 4-beam LiDAR is really sensible .... just makes sense.

The runtime in the paper is reported at 90 ms/frame on a single GPU but the specs of the GPU are not reported.

Would you mind to share which GPU you used to obtain the 90 ms/frame reported in the paper?

You mentioned some CUDA improvements, did you run any experiments with CUDA improvements (not reported in the paper)?

I am curious what would be an acceptable frame rate for a real autonomous driving scenario? A single GPU would get ca. 11 frames/second essentially, have you guys tried multiple GPU's? Or anything similar not reported in the paper to keep things simple?

I re-produced the first PL paper using your AnyNet implementation for disp estimation and saw a dramatic improvement in processing speed with limited (although at times noticeable) difference.

I am looking forward to doing this same again here. Hopefully the L# idea can really make up that difference for AnyNet and still be efficient for embedded platforms

[Jinyong-Huang]

the paper report is just GDC algorithm runs in 90 ms/frame using a single GPU.

[mileyan]

Hi @tareeqav ,

It is good to know that you are using AnyNet on PL framework. As @Jinyong-Huang points, 90ms/frame is the GDC algorithm speed. The whole PL++ pipeline will take more than 500ms per frame. Looking forward to your amazing work.

[Mamoanwar97]

Hello @mileyan and @tareeqav

I am trying to do a similar thing to reproduce a pipeline similar to pseudo lidar repo with Anynet. we now trained the anynet on the kitti object dataset but the resulted point cloud is not clean as either of this pseudo lidar or pseudo lidar ++

Anynet point cloud output:

Psmnet point cloud output:

as you see both are almost the same except for the noise around the car in Anynet point cloud, so do you have any idea how to improve anynet output? @tareeqav did you make any changes to anynet model to fit in the pipeline? or do you know if there is any filtering algorithm that can help in removing this noise?