Preprocessing program

[karljackab]

Hi!
Thanks for your excellent works!

I just wonder that could you provide the preprocessing program? Since I cannot re-implement the MAP score at MSCOCO and NUS-WIDE.

Thanks a lot!

[ymcidence]

Hi!
Thanks for your excellent works!

I just wonder that could you provide the preprocessing program? Since I cannot re-implement the MAP score at MSCOCO and NUS-WIDE.

Thanks a lot!

Hi there

Thank you for your question. Unfortunately, I am no longer with IIAI and the previous data are not getable.

Just for hints, to build tfrecords, please refer to util/data/set_processor.py.

As per feature extractors, simply use CNNs implemented in tf.keras.applications or tensorflow_hub. Make sure the built-in pre-processing function is used.

As per performance, one can slightly tune L18 of layer/twin_bottleneck.py. Alternatively, different sizes of continuous bottleneck and the proportion of regularization would be the help as well.

Note that in the paper we reported map@k scores. Please make sure the evaluation metrics are identical.

Also, note that the map scores on the tensorboard would be lower than the actual value. Please run inference on the WHOLE gallery/query datasets and then test the score after training.

[yanbinbi]

I wonder how you test the score,as i don't see any code for testing