File size: 2,313 Bytes
5e265ac
 
 
 
 
a6403aa
5e265ac
 
 
 
a6403aa
 
5e265ac
 
 
 
a6403aa
5e265ac
 
 
 
 
 
a6403aa
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5e265ac
 
 
 
 
a6403aa
5e265ac
a6403aa
 
5e265ac
 
a6403aa
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
---
library_name: pytorch
tags:
  - humanoid-robotics
  - video-to-motion
  - motion-tracking
---

# ViT-H Checkpoints

This repository contains the released checkpoints used by ViT-H inference and
the optional G1 motion tracker.

## Contents

- `vith_offline_g1.ckpt`: released offline G1 model checkpoint (epoch 629).
- `model_step_020000_g1.onnx`: 1570-input, 29-output G1 tracking policy used by `g1_tracker`.
- `inputs/checkpoints/yolo/yolov8x.pt`: YOLOv8x detector and tracker weights.
- `inputs/checkpoints/vitpose/vitpose-h-multi-coco.pth`: ViTPose-H pose-estimation weights.
- `inputs/checkpoints/hmr2/epoch=10-step=25000.ckpt`: optional HMR2 image-encoder weights.
- `inputs/checkpoints/dpvo/dpvo.pth`: optional DPVO camera-estimation weights.
- `SHA256SUMS`: file integrity checksums.

The SHA256 checksum of `model_step_020000_g1.onnx` is
`da875c31057e4d5f0e0b3ba30722c506fb3a9fa3e030998069b5ded3ac38403b`.

## G1 tracker policy

Download the policy into the ignored checkpoint directory of a
[ViT-H](https://github.com/tany3059/ViT-H-release) checkout:

```bash
hf download eerrr1/ViT-H-checkpoints \
  model_step_020000_g1.onnx \
  --repo-type dataset \
  --local-dir inputs/checkpoints/g1_tracker
```

Then pass it to the tracker:

```bash
python g1_tracker/tools/track_g1.py \
  --input outputs/example/motion.npz \
  --output outputs/example/tracked_motion.npz \
  --input-fps 30 \
  --policy-onnx inputs/checkpoints/g1_tracker/model_step_020000_g1.onnx
```

See the [G1 tracker documentation](https://github.com/tany3059/ViT-H-release/tree/main/g1_tracker)
for the full command, robot assets, rendering options, and output format.

## Other dependencies

The released configuration uses SAM3D image features. HMR2 is included as an
optional compatible image encoder. DPVO is only required when inference uses
`--use-dpvo`; SimpleVO remains the default camera-motion estimator.

SAM3D code and weights are not included in this checkpoint repository. Install
[Fast-SAM-3D-Body](https://github.com/yangtiming/Fast-SAM-3D-Body) under
`third-party/Fast-SAM-3D-Body` and obtain its weights through the upstream
project. Install optional [DPVO](https://github.com/princeton-vl/DPVO) under
`third-party/DPVO`.

External model files remain subject to their respective upstream licenses and
access terms.