SLiM

[NeurIPS 2026] SLiM

Encoders from Less is More: Compact-Token Masked Feature Learning for Skeleton Representation Learning.

• Paper: https://arxiv.org/abs/2603.10648
• Project page: https://kaist-viclab.github.io/SLiM_site/
• Code: https://github.com/KAIST-VICLab/SLiM

SLiM patchifies a 64-frame skeleton clip with P_T = 8, P_J = 1 into a compact 8 x 25 token grid and trains a ViT encoder (8 blocks, dim 256, 8 heads) against an EMA teacher with masked feature prediction and global-local contrastive learning. No coordinate decoder is used, and downstream inference costs 3.59 GFLOPs.

Available checkpoints

One encoder per evaluation protocol, each in its own subfolder. Linear evaluation uses a frozen encoder, a linear classifier and a single joint stream.

NTU RGB+D 120 weights are coming and will be added here as ntu120_xsub and ntu120_xset.

Usage

git clone https://github.com/KAIST-VICLab/SLiM && cd SLiM
pip install -r requirements.txt huggingface_hub safetensors

xFormers and a CUDA GPU are required — the temporal-RoPE attention has no CPU path.

import torch
from slim.hub import SLiMEncoder

model = SLiMEncoder.from_pretrained("JeonghyeokDo/SLiM", subfolder="ntu60_xsub").eval().cuda()

# clips: (B, 3, 64, 25, M) — coordinates, frames, joints, people
clips = torch.zeros(2, 3, 64, 25, 1).cuda()
with torch.no_grad():
    feats = model.get_intermediate_layers(clips, 4, return_class_token=True)
patch_tokens, cls_token = feats[-1]   # (B, 200, 256), (B, 256)

Evaluate each encoder on its own protocol only. Inputs follow the repository's preprocessing: clips are resampled to 64 frames, one person per token sequence, coordinates shifted so that joint 1 of the first frame is the origin.

Files

ntu60_xsub/
├── config.json, model.safetensors      encoder — load with subfolder="ntu60_xsub"
└── linear/                             the trained linear-probe grid behind the number above
ntu60_xview/
├── config.json, model.safetensors      encoder — load with subfolder="ntu60_xview"
└── linear/

See the repository README for the evaluation commands.

Citation

@inproceedings{do2026slim,
  title     = {Less is More: Compact-Token Masked Feature Learning for Skeleton Representation Learning},
  author    = {Do, Jeonghyeok and Chen, Yun and Youk, Geunhyuk and Kim, Munchurl},
  booktitle = {Advances in Neural Information Processing Systems (NeurIPS)},
  year      = {2026}
}

Built on DINOv2 (Apache 2.0).