Request for Pre-processed Data Sample for Virtual Try-on Demo #3

Activity

1. azuxmioy commented on Apr 23, 2026

   

   azuxmioy

   on Apr 23, 2026

   Contributor

   More actions

   Hi Vikas,

   Thanks for reaching out. We'll take a look and keep you posted here

   @zhiyi-chen217
2. zhiyi-chen217 commented on May 6, 2026

   

   zhiyi-chen217

   on May 6, 2026

   Collaborator

   More actions

   Hi Vikas,

   Thanks for the comment!

   • I have added the checkpoints in the README.
   • Here is a small template you can start from for the exchange cloth data.
     • Most of the files needed are already there. You just need to generate the pose maps
     • Here I also included a sequence of poses you can try "00127_train.npz"
     • Here, I included a 134_140.yaml file you can use as the config, you can tweak with the smpl_pos_map, sequence_name, exchange_cloth->change_mode and checkpoints to experiment with different configurations

   👍React with 👍1Reacted by WU YUZE
3. wuyuze713 commented on Oct 15, 2026

   

   wuyuze713

   on Oct 15, 2026

   More actions

   Hi @zhiyi-chen217, thank you for sharing the checkpoints and exchange example. I would like to reproduce the original 00127 Inner input conditions before changing the method.

   At commit af9ff485602b55f8125651e48803b1448b67c0b8, configs/train_test_division.md lists training Takes 2/4/6/7/8 and test Takes 9/10. The shared body/00127_train.npz contains 735 pose rows, consistent with the combined frame counts of those five training Takes. The shared body/cameras.pkl contains the 12 camera IDs selected by the training configuration.

   However, the public 00127 archives contain Capture images/masks only for 0004, 0028, 0052, 0076. Could you please clarify:

   1. Where can the remaining training views (0016, 0036, 0041, 0048, 0060, 0069, 0084, 0092) and their matching masks/calibration be obtained, subject to the relevant data permissions? Are these additional views captured RGB, or is there a documented rendering/preparation step?
   2. Does the cameras.pkl in the 00134-clothing/00140-body exchange example also apply to 00127? If not, is the corresponding rgb_cameras.json or prepared camera file available?
   3. Is there a complete 00127 preprocessing example or the referenced local meshavatar template-merging/body-mesh-generation code? We want to match the template and body supervision, not silently substitute different geometry.
   4. The public test Takes contain 140 + 150 = 290 frames, whereas the paper describes 300 test frames per subject. Is 300 approximate, or are additional frames used? A prepared-index-to-Take/frame manifest (including any excluded frames) would help reproduce the split exactly.

   Pointers to existing resources would already be very helpful. Thank you!
4. zhiyi-chen217 commented on Oct 17, 2026

   

   zhiyi-chen217

   on Oct 17, 2026

   Collaborator

   More actions

   Hi @wuyuze713,

   Thank you for your detailed questions:

   1. Sadly we cannot provide the remaining training views. These addtional views are captured RGB, similar to the public training views. If needed, we can provide checkpoints for subject 00127

   2. Yes, the cameras.pkl applies to all 4D-Dress subjects

   3. Unfortunately, we cannot provide the exact templates we are using. However, the scripts are available. You can use gen_data/create_mesh_avatar.py to generate the canonical template and use gen_data/cano_to_shaped.py to restore the canonical templates to shape-zero space. The needed assets should be provided by the 4D-Dress.

   4. Yes, 300 is an approximation