Ground Truth Data Requirement for Training a Point Piller Network on Lidar Data?

Hi Everyone :) First of all thank you to everyone of you out there helping newbies like me.

I want to train a Point Piller Network with Lidar Point Clouds. However, I am not sure, how much data should I label before, using Lidar Labeler app to generate ground truth. I have a data from Velodyne Lidar mounted on the top of vehicle. Could you please tell me, how much labelled data (number of frames) shall be enough? Shall I label every object in a single frame? or one label per frame could also work in this case?

My ultimate goal then, is to use this preTrained detector in Lidar Labeling Algorithm to ground Truth even a larger dataset.

Thank you very much.

Answer 1

[Vinay Ancha]

Hi,

There is not hard and fast rules about how much data you should train for a network to perform better. As you mention that the requirement is train a subset of dataset to get a pretrained model and use it to annotate the remaining data, I would suggest you to randomly select around 30-40% of the dataset and annotate it. If the dataset contains continous sequence, do not select continous frames for training. Once you get a pretrained model try and see how it performs on the remaining dataset. If you think that it still needs improvement you can annotate another 10% of the dataset, train it and check the performance. Annotate every object in a single frame as during training it will consider the other objects (of interest) that aren't annotated as background and it would affect the network accuracy.

[Sarang Jokhio]

Hi Vinay,

Thank you so much for the response :)

Could you please explain this further?

"If the dataset contains continous sequence, do not select continous frames for training".

Do you mean that among the 40% data, that I select for training, I annoate random frames?

Sorry if my question does not make sense :D I am super new into this.

Thank you

[Vinay Ancha]

Hi Sarang,

Do you mean that among the 40% data, that I select for training, I annoate random frames?

No, you should annotate the full 40% data. If your dataset contains continous sequences, select the dataset for annotation such that the selected sequences do not contain continous frames.