data_doc.md

February 6, 2026 ยท View on GitHub

Data documentation

Data download

After registering an account at here and then you will recieve a link to download PALM-Net/PALM and have access to the following files:

./PALM/XR20A/zips/0117.zip
./PALM/XR20A/zips/0201.zip
./PALM/XR20A/zips/0140.zip
./PALM/XR20A/zips/0251.zip
...
./PALM/XR20A/quality_dict_feb_27.npy
./PALM/XR20A/scan_ids.txt
./PALM/XR20A/MCU_04_env.exr
./PALM/XR20A/meshes/0221.zip
./PALM/XR20B/zips/0221.zip
./PALM/XR20B/zips/0226.zip
./PALM/XR20B/zips/0251.zip
./PALM/XR20B/zips/0003.zip
./PALM/XR20B/zips/0049.zip

...
./PALM/XR20B/quality_dict_feb_27.npy
./priors/0301-1442-124.zip
./pca_basis/0301-1442-124/epoch=0-step=24000.appearance.pca.npy
./pca_basis/0301-1442-124/epoch=0-step=24000.shape.pca.npy
./pretrained/flat_hand.right.dim32.7subj.feb.23.pt
./personalize_ckpts/0306-0218-596.zip
./personalize_ckpts/0306-1503-223.zip
./gt_checksums.json

There are two versions of PALM: XR20A and XR20B. The XR20A (137GB) version contains high-resolution 2K images while the XR20B (27GB) version resized the images for faster training. The camera intrinsics for XR20B are resized accordingly. Everything else in these two versions are the identical. We also have 3dMD scans for each frame at PALM. All 3D annotations are in meters.

  • ./PALM/XR20A/zips/0117.zip: subject 0117 of PALM dataset with version XR20A.
  • ./PALM/XR20A/quality_dict_feb_27.npy: fitting error for each frame.
  • ./PALM/XR20A/MCU_04_env.exr: calibrated environment map
  • ./PALM/XR20A/meshes/0221.zip: Hand scans for subject 221
  • ./priors/0301-1442-124.zip: Pretrained multi-subject prior with version 0301-1442-124 (see Checkpoints below)
  • ./pca_basis/0301-1442-124/epoch=0-step=24000.appearance.pca.npy: PCA appearance space of the prior with version 0301-1442-124
  • ./pca_basis/0301-1442-124/epoch=0-step=24000.shape.pca.npy: PCA shape space of the prior with version 0301-1442-124
  • ./pretrained/flat_hand.right.dim32.7subj.feb.23.pt: pretrained geometry network on 7 subjects using MANO SDF values; we found using a pretrained hand geometry network stablizes the PALM model training.
  • ./personalize_ckpts/0306-0218-596.zip: Personalization checkpoints in our InterHand2.6M experiments (see Checkpoints below)
  • ./gt_checksums.json: checksums of the files to check for file corruption.

About the calibrated environment map, we put a metal sphere in our 3dMD capture setup and fit a sphere to this metal sphere to unwrap an environment map from RGB observations.

Checkpoints

We offer three versions of priors trained on 14%, 50% and 100% of frames at PALM:

PriorsNumber of frames
0301-1442-12414% (ckpt in paper)
0311-0806-12450%
0311-0910-389100%

Here are the personalized checkpoints for our InterHand2.6M experiments: ih_c0_ROM03_RT_No_Occlusion_cam400262 denotes the sequence with name ROM03_RT_No_Occlusion in capture 0 on the test set; cam400262 is the camera view.

InterHand2.6M sequencesPersonalized checkpoints
ih_c0_ROM03_RT_No_Occlusion_cam4002620306-0218-303
ih_c0_ROM03_RT_No_Occlusion_cam4004510306-0218-212
ih_c0_ROM04_RT_Occlusion_cam4002750306-0218-171
ih_c0_ROM04_RT_Occlusion_cam4004180306-0218-458
ih_c0_ROM05_RT_Wrist_ROM_cam4002700306-0218-299
ih_c0_ROM05_RT_Wrist_ROM_cam4004880306-0218-402
ih_c1_ROM03_RT_No_Occlusion_cam4004560306-0218-443
ih_c1_ROM03_RT_No_Occlusion_cam4004860306-0219-214
ih_c1_ROM04_RT_Occlusion_cam4002660306-0218-596
ih_c1_ROM04_RT_Occlusion_cam4004390306-0219-262
ih_c1_ROM05_RT_Wrist_ROM_cam4003140306-0218-439
ih_c1_ROM05_RT_Wrist_ROM_cam4004690306-0218-495

Environment maps

Environment maps can be found here. You should put them under code/envs/*.hdri and code/envs/*.exr.