search

maria-korosteleva / Garment-Pattern-Estimation

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

Bad stitch prediction results #3

Closed zycliao closed 11 months ago

zycliao commented 1 year ago

Hi, thanks for the amazing work. I'm trying to test the pre-trained model on the official dataset. However, the result looks bad. I'm wondering if I did anything wrong.

Basically, I used the pre-trained weights and tested on a few garments provided by the dataset. I ran predict_per_example.py with the default configurations. For example, when I tested garment_pattern/dress_150/dress_2Q19HWLQ83/dress_2Q19HWLQ83_sim.obj, the result was as below, image but after importing it into Garment Viewer in Maya, the result was image

maria-korosteleva commented 1 year ago

Hi @zycliao, thank you for your interest in our work!

  1. Could you specify what exactly is not meeting your expectations? The shape of the panels is close to GT, and they are placed +/- correctly relative to the body. Stitches are quite bad, but, as reported in the paper, the stitch prediction on unseen types is, unfortunately, not ideal:

    image

    So, I would say that the presented example would match my expectations 🤔

    For reference, this is the GT pattern from a garment in question dress_2Q19HWLQ83_pattern

  2. To double-check if your setup is correct, may I ask you to run the evaluation script to see if the accuracy numbers on your installation are similar to the ones reported in the paper? The command is as follows, running from the directory root:

    python nn/evaluation_scripts/on_test_set.py -sh models/att/att.yaml -st models/att/stitch_model.yaml --unseen

    Please, report the numbers that you will get

zycliao commented 1 year ago

@maria-korosteleva Thanks for your reply. I was mainly concerned about the stitch (the blue lines in my screenshot). They look totally different from GT. And after the simulation, it becomes a mess.

maria-korosteleva commented 1 year ago

Yes, please, report the results on stitches accuracy from the evaluation script, as I requested above, to double-check that the setup is correct

If you have a need to use this specific prediction in the downstream application, you can adjust (add/remove) the stitches manually following Qualoth's manual