Straighten out RWTH-PHOENIX-* citations.

cleong110 commented 1 week ago

There are a number of places where citations or confusions might need to be double-checked.

Search for "PHOENIX" and confirm we're citing the right ones.

forster-etal-2012-rwth aka RWTH-PHOENIX-Weather introduction of the dataset.
dataset:forster2014extension extends it, that's RWTH-PHOENIX-Weather 2014
cihan2018neural makes it a translation set, this is where RWTH-PHOENIX-Weather 2014T comes from
Continuous sign language recognition: Towards large vocabulary statistical recognition systems handling multiple signers aka koller2015ContinuousSLR

See also #70, no idea why I thought koller2015ContinuousSLR was relevant, possibly some other paper cited it incorrectly?

cleong110 commented 1 week ago

For example we have this in SignBERT+

continuous SLR (RWTH-PHOENIX-Weather [@koller2015ContinuousSLR]

In the SignBERT+ paper they cite thus:

For continuous SLR, the evaluation is conducted on two datasets, i.e., RWTH-Phoenix [14] and RWTH-PhoenixT [33]

and 14 is

O. Koller, J. Forster, and H. Ney, “Continuous sign language recognition: Towards large vocabulary statistical recognition systems handling multiple signers,” Computer Vision and Image Understanding (CVIU), vol. 141, pp. 108–125, 2015.

But that's the wrong citation! That citation is not the source of any dataset. It does have analysis...

In this paper, we present extensive results and thorough analysis on, to our knowledge, the currently biggest publicly available corpus of continuous SL (RWTH-PHOENIX-Weather)

and

RWTH-PHOENIX-Weather database 9 signer, 1081 sign vocabulary, 7k sentences

(which is again wrong, the 9 signer version is RWTH-PHOENIX-Weather 2014)

There's also this:

So which one did SignBERT+ use?

cleong110 commented 1 week ago

SignBERT+ says

For continuous SLR, the evaluation is conducted on two datasets, i.e., RWTH-Phoenix [14] and RWTH-PhoenixT [33]. RWTH-Phoenix [14] is a popular German sign language dataset collected from the weather forecast broadcast. It contains 6,841 samples, with 5,672, 540 and 629 videos for training, validation and testing, respectively.

OK, we can look for those numbers: