How to use background data with Lite Model Maker Object Detection?

General Discussion
, ,
#1

Hello, I am trying to setup an object detection model similar to the one described here

The model works fine with labeled data, but when I try to use background data that features no bounding boxes in the CSV file, I get the following error.

Traceback (most recent call last):
File “makemodel.py”, line 20, in
train_data, validation_data, test_data = object_detector.DataLoader.from_csv(filename=os.getcwd() + ‘/csvfile.csv’, )
File “/home/m/.local/lib/python3.6/site-packages/tensorflow_examples/lite/model_maker/core/data_util/object_detector_dataloader.py”, line 292, in from_csv
cache_writer.write_files(cache_files, csv_lines=csv_lines)
File “/home/m/.local/lib/python3.6/site-packages/tensorflow_examples/lite/model_maker/core/data_util/object_detector_dataloader_util.py”, line 247, in write_files
for idx, xml_dict in enumerate(self._get_xml_dict(*args, **kwargs)):
File “/home/m/.local/lib/python3.6/site-packages/tensorflow_examples/lite/model_maker/core/data_util/object_detector_dataloader_util.py”, line 381, in _get_xml_dict
lines)
File “/home/m/.local/lib/python3.6/site-packages/tensorflow_examples/lite/model_maker/core/data_util/object_detector_dataloader_util.py”, line 337, in _get_xml_dict_from_csv_lines
xmin, ymin = float(line[3]) * width, float(line[4]) * height
ValueError: could not convert string to float:

It seems to be treating the CSV lines containing background data as it would treat a CSV line containing a bounding box, and I am not sure why it is not recognizing the CSV line as a line of background data. I followed the formatting guidelines from here. Formatting a training data CSV  |  AutoML Vision Object Detection  |  Google Cloud.
Any ideas on what could be causing this?

#5

Thank you for the reply. However, I’m still not sure how I format the background picture data so that the model accepts it, and I would appreciate any help on this topic.

#7

I have the same problem.
According to the documentation they should have a csv line with none values for xmin, ymin, etc.
The source could looks like that though:

for line in lines:
    label = line[2].strip()
    xmin, ymin = float(line[3]) * width, float(line[4]) * height
    xmax, ymax = float(line[7]) * width, float(line[8]) * height

This obviously does not work if the columns are just empty. Furthermore, there is also this line:

self.label_name2id_dict = {'background': 0}

Which is similar to the Object Detection API. But when i use “background, 0,0,0,0” in the csv file (which was the correct way in the OD API), the label background still gets an id that is not 0.