search

borutb-fri / FMLD

A challenging, in the wild dataset for experimentation with face masks with 63,072 face images.
MIT License
36 stars 7 forks source link

I think theres some bad labels in the annotations zip. bbox coords are outside image dimensions. causes error in tensorflow #2

Open late347 opened 3 years ago

late347 commented 3 years ago

issue 1: These was not necessarily the only bad data I found. But it was what I found from my dataset when I inspected the csv file, after having gotten crash from illegal bounding box outside image dimensions.

File "C:\Users\lauri\anaconda3\envs\tf_kaggle\lib\site-packages\tensorflow\python\ops\control_flow_ops.py", line 154, in Assert raise errors.InvalidArgumentError( tensorflow.python.framework.errors_impl.InvalidArgumentError: Expected 'tf.Tensor(False, shape=(), dtype=bool)' to be true. Summarized data: b'maximum box coordinate value is larger than 1.100000: ' 1.1067415

The ones I found from my subset of your FMLD dataset with sketchy bbox coordinates were these in csv

filename test_00000749.jpg test_00001127.jpg test_00001586.jpg test_00001607.jpg test_00001626.jpg test_00001626.jpg test_00003030.jpg test_00003055.jpg test_00003672.jpg test_00003750.jpg test_00003774.jpg test_00004007.jpg test_00004179.jpg   filename train_00000019.jpg

INVALID_COORDINATES_excelsheet_test_train.xlsx

Issue 2: I found Hitler the dataset. I think he was maybe hiding in the WIDERFACE. picture name 0_Parade_Parade_0_914.jpg But I managed to delete him from my dataset as well :)

borutb-fri commented 2 years ago

Thanks for the warning. There is no bad labels in the annotations in the data. Only the coordinates are written as returned by the face detector.

You can very easily limit the coordinates to the allowed values, as I did in the code: show_save_gt.m in line: 53-56

xmin = max (1, xmin); ymin = max (1, ymin); xmax = min (width, xmax); ymax = min (hight, ymax);

twmht commented 2 years ago

@borutb-fri

Did you remove some of the bounding boxes from widerface? It seems that some of the images have missing bounding boxes.

borutb-fri commented 2 years ago

I use only the subset from the Wider Face dataset. For original bounding boxes of widerface refer to WIDERFACE.

twmht commented 2 years ago

@borutb-fri

Have you ever compared using image classifier and object detector on the same training dataset?