Person-in-WiFi 3D — Pretrained Checkpoints

Re-implementation of Person-in-WiFi 3D (CVPR 2024).
Estimates 3D human body keypoints from Wi-Fi Channel State Information (CSI) — no camera required.

Code: github.com/Duongvu05/hpe_task

---

Files

---

Training Setup

---

Results (best.ckpt)

Overall by person count (mm):

Per-joint breakdown (mm):

---

Usage

1. Download checkpoints

hf sync hf://buckets/VuNgocDuong/Hpe-task ./work_dirs

2. Clone repo & install

git clone https://github.com/Duongvu05/hpe_task
cd hpe_task
uv sync

3. Evaluate

./evaluate.sh \
  --checkpoint work_dirs/best.ckpt \
  --train-root train_data \
  --test-root  test_data

4. Resume training

./train.sh --resume-from work_dirs/best.ckpt

---

Model Architecture

Wi-Fi CSI  [B, 3, 3, 20, 60]
     │
     ▼  Linear projection + STE positional encoding
     │
  Encoder  (6 × Transformer, 8 heads, dim=256)
     │
     ▼  Two-stage proposals (top-100 encoder outputs)
     │
  Decoder  (3 × Transformer)  — iterative keypoint refinement
     │
     ▼  Hungarian matching → positive instances only
     │
  Refine Decoder  (3 × Transformer)
     │
     ▼
  Output: [B, 100, 14, 3]  (14 keypoints × xyz, mm)

---

Reference

@inproceedings{yan2024personinwifi3d,
  title     = {Person-in-WiFi 3D: End-to-End Multi-Person 3D Pose Estimation with Wi-Fi},
  author    = {Yan, Xinyan and Wang, Xu},
  booktitle = {CVPR},
  year      = {2024}
}