Inference and Training about lunet HOT 1 CLOSED

Hi,

Q1. You can use the inference on a custom dataset. You just need to generate a TFRecords in the same way as it is done during the data preparation.
Q2. LU-Net has only been tested on the SqueezeSeg dataset (created from KITTI) and its 3 classes "car", "pedestrian", "cyclist". However it is possible to train it on different datasets such as SemanticKITTI to be able to distinguish more classes.
Q3. If you want to use the code as is, the custom dataset has to match the standards of the SqueezeSeg dataset, meaning that each sample should be a 64x512x6 matrix whose channels are respectively: x, y, z, depth, reflectance, label
Q4. I do not plan on sharing the pretrained model, but you can follow the training protocol of the README to retrain the model, which takes several hours on a good GPU, and achieve similar results as in the paper.
Q5. I'm not sure I understand the question, the performances of the model are described in the associated publication. The validation scores are computed on real data.

Hope it helps :)

from lunet.