Request a simple demo script

[peiyunh]

Hi,

Thanks for sharing your amazing work.

This is a shot in the dark, but I am wondering if you consider releasing a simple demo script that takes a LiDAR sweep as input and output predicted semantic labels. I am asking because I want to run the pretrained models on a different dataset. The easiest way seems to be porting my dataset into the same format as SemanticKITTI. Is that the case?

Thanks,
Peiyun

[tano297]

Hi,

Not a shot in the dark at all.

I thought of providing this functionality as I do in bonnetal with an infer_video.py and infer_img.py, but unfortunately, images are a lot more standard than point clouds to deal with.

So, I quickly realized that to make an infer_scan.py script I would have to specify a format, and if we decided on our own format, then our inference script was already good enough.

I could provide a script, for example, for .ply, .stl, or .pcd files, or something standard like that, but since I am going to release a ROS interface soon after ICRA deadline I let this side project go for now.

So, long story short, putting the data in the kitti format is the easiest way for now.

[peiyunh]

Thanks for the reply! Closing the issue.

[ivannson]

@peiyunh Were you able to run get the predictions for your own dataset? And did you get those predictions after training with KITTI data?

@tano297 I noticed that in infer.py script, the data split into train, val and test, so there has to be at least one sequence from 00 to 07 for train, 08 for val, and any other sequence for the test. I was wondering what is the reason for doing so and if it's possible to skip train and val, since the model is already trained, and go directly to predictions?

[tano297]

@ivannson. It is completely possible.

2 ways to do this:

• Simplest: Just split the test set into the 3 categories again, i.e:

Changing

split: # sequence numbers
  train:
    - 0
    - 1
    - 2
    - 3
    - 4
    - 5
    - 6
    - 7
    - 9
    - 10
  valid:
    - 8
  test:
    - 11
    - 12
    - 13
    - 14
    - 15
    - 16
    - 17
    - 18
    - 19
    - 20
    - 21

for

split: # sequence numbers
  train:
    - 11
  valid:
    - 12
  test:
    - 13
    - 14
    - 15
    - 16
    - 17
    - 18
    - 19
    - 20
    - 21

• Another way is by going to modules/user.py and commenting the following lines:

  def infer(self):
    # do train set
    # self.infer_subset(loader=self.parser.get_train_set(),
    #                 to_orig_fn=self.parser.to_original)

    # do valid set
    # self.infer_subset(loader=self.parser.get_valid_set(),
    #                  to_orig_fn=self.parser.to_original)
    
    # do test set
    self.infer_subset(loader=self.parser.get_test_set(),
                      to_orig_fn=self.parser.to_original)

    print('Finished Infering')

    return

Both are hacks. Feel free to modify the infer function to take something like "do_train, do_valid, do_test" as parameters and submit a PR, I'm happy to change this

[ivannson]

Great, will do that. Thanks a lot for such a quick reply!

[peiyunh]

@ivannson Yes, I was able to get the predictions on my own dataset. I basically took the code from infer_subset and replaced the existing data loader with an iterator over my own dataset.

[abhigoku10]

@peiyunh is it possible ot share the inference pipeline on the custom dataset so that i can try it with my own data ?? thansk in advance

[xdtzzz]

@peiyunh hi, i also do this work, but i don't know how to use infer.py to get predictions with my own dataset. Where should to be modified? i only test my own dataset, what are the train, valid and test dataset?