pair-code / knowyourdata Goto Github PK

Bounding boxes for the kitti dataset are vertically flipped. This is a bug with the TensorFlow datasets implementation.

See: https://knowyourdata-tfds.withgoogle.google.com/#dataset=kitti&tab=ITEM&draw=kyd/kitti/objects_type,bbox,bbox&item=000079.png

See the BBoxFeature documentation on TensorFlow datasets: https://www.tensorflow.org/datasets/api_docs/python/tfds/features/BBoxFeature

This states:

bbox: tf.Tensor of type tf.float32 and shape [4,] which contains the normalized coordinates of the bounding box [ymin, xmin, ymax, xmax]

Really nice demo!

I'm curious how exposure_quality and sharpness_score are computed. Both seem somewhat non-trivial to measure and I'm not sure what the standard process is in ML (as opposed to human perception).

I believe would be a great addition to the documentation if you add a proper citation format for the too.
Is there a specific whitepaper? Should we reference People + AI Research, Google, or Both?

Just like the guidelines in Tensorflow:
https://www.tensorflow.org/about/bib

I have some remote sensing images, and it would be helpful to use this structure to insert some samples/classes to teach deep learning for our applications.
Regards

I don't see any textual data in the demo. Are there any plans to have text data?

Hello there, when are we going to support custom dataset?

Hi folks.

Is there any plan to set Know Your Data locally and run it on custom datasets? I think if the users could have their datasets in a way TensorFlow Dataset expects it might be doable.

looks like it can't make toast

KYD mentions accepting Creative Commons licensed datasets, but the VGGFace2 dataset is actually comprised of copyright photos. Many even include watermarks over the image with the names of the photo agency.

Here are a few photos that even include the watermark:

https://knowyourdata-tfds.withgoogle.com/#dataset=vgg_face2&tab=ITEM&select=kyd%2Fvgg_face2%2Flabel&item=n000002%2F0034_01.jpg
https://knowyourdata-tfds.withgoogle.com/#dataset=vgg_face2&tab=ITEM&select=kyd%2Fvgg_face2%2Flabel&item=n000002%2F0072_02.jpg
https://knowyourdata-tfds.withgoogle.com/#dataset=vgg_face2&filters=kyd%2Fvgg_face2%2Flabel:n000026&tab=ITEM&select=kyd%2Fvgg_face2%2Flabel&item=n000026%2F0116_01.jpg
https://knowyourdata-tfds.withgoogle.com/#dataset=vgg_face2&filters=kyd%2Fvgg_face2%2Flabel:n000026&tab=ITEM&select=kyd%2Fvgg_face2%2Flabel&item=n000026%2F0116_01.jpg
https://knowyourdata-tfds.withgoogle.com/#dataset=vgg_face2&tab=ITEM&select=kyd%2Fvgg_face2%2Flabel&item=n000006%2F0020_01.jpg

If you keep scrolling through it you'll notice hundreds, maybe thousands, others from Alamy, Shutterstock.

The only component that is licensed with a "Creative Commons" is the list of images, not the actual images. It's not possible to override copyright with Creative Commons.

Hello I am wondering if there is any starter task posted for this project?

Attention, the i_naturalist2017 dataset has the labels misaligned from the items to which they refer, therefore, in the current state, it is completely useless.

How can I locally host this tool to explore local datasets? Using local.sh simply redirects to https://knowyourdata-tfds.withgoogle.com