Question regarding input and output tensors of Pointpillars

[aatefi2]

Hi shangjie-li,

I hope you are doing well. Thank you for all of your help since every time I can not find the exact answer from searching on different webpages, you give me very useful suggestions.

My questions:

1) If I just have pointcloud data (without 2D images and calibration files), I am wondering what is the input and output tensors (data type, shape, name, ...) of Pointpillars for test dataset (test.py and demo.py)? Is it possible to print them from the provided codes on your Github?

2) I found the information below:

• Input tensor: from the Pointpillars paper and other webpages, I found the input tensor is [D, P, N] which D=9 (I guess 9 float numbers, which float type?), P=number of pillars (defined in cfg.yaml and is 40000 for testing data), N= the number of points per pillar (defined in cfg.yaml and is 32 for testing data). I think we can assume P,N as int values (which int type?).
• Output tensor: Regarding the output tensor, I could get three different output tensors: 1) predicted box [M,7], 2) score [M], and 3) class name [M] which M is the number of the predicted objects. I think the predicted box and score are float (which float type?), and I found the class name as int64 after running demo.py. Please find the snapshots below for 0,1, and 3 predicted objects.

Please let me know whether this information that I found is valid or not.

Thank you, Abbas

[shangjie-li]

I am very happy to know that my advices can help you. I am usually busy with work, but I will try my best to answer your questions.

The implementation of this code is basically the same as the original Pointpillar paper, but there are some differences in the details. If you want to get a deeper understanding of how the network works and how the dimensions of the tensors change at each step, I suggest you read the code in the layers directory in detail. In addition, pointpillar.py organizes the modules in the layers together to form a completed inference process. I find this process difficult to convey in short words. And, from what I understand, the process of tensor dimensional transformation is the most important part of network inference.

For a tensor a, if you want to see its information, you can add some debug code to the source code. For example, print(a.shape) is used to see the dimensional size of the tensor, and print(a.dtype) is used to see the data type of the tensor.

[aatefi2]

Hi shangjie-li,

Thank you for your useful suggestions.