search

maria-korosteleva / Garment-Pattern-Estimation

Reconstructing garment sewing patterns from 3D point clouds
MIT License
112 stars 15 forks source link

Selection of the panel decoder architecture. #2

Closed chenghsiuchen closed 1 year ago

chenghsiuchen commented 1 year ago

Hi,

Your work is amazing and thank you for sharing the code and dataset. I am curious about why do you select "LSTMDecoderModule" over other candidates (e.g., "MLPDecoder") as the panel decoder. Can you share some insights about the performance difference or artifacts when using alternative decoders ? It seems that this wasn't discussed in the paper, but alternative modules do exist in the code.

Thanks in advance.

maria-korosteleva commented 1 year ago

Hi! Thank you for your kind words and interest in our work!

We originally chose LSTM for our panel decoders, as the panels have a natural sequential representation, and LSTMs allowed weight sharing and flexibility with sequence length. When combined with the Loop Loss, the panel predictions became high enough in quality, so we didn't have much reason to change that part of the architecture.

We did some experiments with MLPDecoder to confirm our choice, but they were done very roughly (e.g. smaller #of garment types in training data). Unfortunately, it's been quite a while since we worked on them, so the only number I could find right now is from an experiment where both pattern and panel decoders were substituted with MLPs. Those were about 40-45 percentage points below our final model (60% vs 95% at the same stage of training) on the number of panels and number of edges accuracies. We decided not to invest further into clearing those experiments up, so they didn't make it into the paper. I left it in the code for other curious minds =)

If you ever decide to experiment with those modules, please, do let us know the outcomes. That would be greatly appreciated

пт, 9 дек. 2022 г. в 20:05, chenghsiuchen @.***>:

Hi,

Your work is amazing and thank you for sharing the code and dataset. I am curious about why do you select "LSTMDecoderModule" over other candidates (e.g., "MLPDecoder") as the panel decoder. Can you share some insights about the performance difference or artifacts when using alternative decoders ? It seems that this wasn't discussed in the paper, but alternative modules do exist in the code.

Thanks in advance.

— Reply to this email directly, view it on GitHub https://github.com/maria-korosteleva/Garment-Pattern-Estimation/issues/2, or unsubscribe https://github.com/notifications/unsubscribe-auth/ABUFLAY55SHG36GWDYIGAD3WMN7JDANCNFSM6AAAAAASZVCJR4 . You are receiving this because you are subscribed to this thread.Message ID: @.***>

chenghsiuchen commented 1 year ago

Thank you very much for your explanation !