regarding the metadata_path

[anguoyang]

Hi @ZhenglinZhou , there is a para named with metadata_path, where can I get it?

python evaluate.py --device_ids=0 \
                   --model_path=${model_path} --metadata_path=${metadata_path} \
                   --image_dir=${image_dir} --data_definition={WFLW, 300W, COFW} \ 

btw, I downloaded the 300w dataset, it seems the annotation of the test data included in test.ts, just wonder how you split it into full, common and challenge and evaluate them respectively? it seems there is no item in the annotation to tell which image is challenge, thank you

[ZhenglinZhou]

Hi @anguoyang!

You can extract the subset of 300W and COFW from the full TSV files by using the image file name provided in the official split files.

If you have any question, feel free to leave a comment.

[anguoyang]

@ZhenglinZhou thanks, understand, but the evaluation result on 300w-full is 0.0287 with the evaluate.py, your report is 2.87, which is 100 times, is it correct?

[ZhenglinZhou]

@anguoyang Yes. We follow the previous research that report NME in percentage format, indicating that the result should be 2.87% = 0.0287 * 100%.