custom object detection dataset issues

[avber]

Hi @leondgarse ,

I've converted coco format dataset to this format (#52 (comment)).

There is only 1 class, so all indices are "label": [1, 1, 1, 1]

Is it correct to specify indices_2_labels in the following format?

"indices_2_labels": {"1": "tp-other"}, "info": {"num_classes": 1}

When I'm checking the dataset, all bboxes are labeled as Person.

tt = data.init_dataset('data.json', batch_size=2, anchors_mode="anchor_free")[0]
data.show_batch_sample(tt, anchors_mode="anchor_free")

The other issue is there are blank boxes when visualizing images. When converting the dataset, I checked that all files exist.

Thanks

[leondgarse]

• The "indices_2_labels" and "info" should be correct.
• The person label is normal. As show_batch_sample not receiving a indices_2_labels parameter, the label name is decoded from COCO, and index 0 is person. Will fix this...
• The empty image is not right. It seems the second image is not plotted, as the axes indices 0.0 ... 1.0 are not turned off, and ax.set_axis_off() cannot be skipped during show_image_with_bboxes... You may try

  images, labels = tt.as_numpy_iterator().next()
  data.show_batch_sample([images, labels], anchors_mode="anchor_free")

  If plotted image is empty, check if labels has correct shape. It shouldn't happen... Did it print any other error?

[avber]

attached is the ds conversion log. Number of labels is equal to number of bboxes for every image.

tt = data.init_dataset('data.json', batch_size=2, anchors_mode="anchor_free", max_labels_per_image=200)[0]

Every time I run these lines, there is an empty box in the top right corner. max_labels_per_image didn't help

images, labels = tt.as_numpy_iterator().next()
data.show_batch_sample([images, labels], anchors_mode="anchor_free")

There are no errors printed.

ds conversion log.txt

[avber]

labels.shape (2, 200, 6)

[leondgarse]

Uh, are you using jupyter notebook plotting that? If yes, the first image with empty box doesn't matter, just an image displaying issue caused by plt.show()...

[avber]

Yeah, it's jupyter notebook