Motion In-betweening via Two-stage Transformers

Jia Qin, Youyi Zheng, and Kun Zhou. 2022. Motion In-betweening via Two-stage Transformers. ACM Trans. Graph. 41, 6, Article 184 (December 2022), 16 pages. https://doi.org/10.1145/3550454.3555454

开始

1. 下载 LAFAN1 数据集.

2. 解压 lafan1.zip 到 datasets 文件夹. Bvh 文件应放在 motion_inbetweening/datasets/lafan1 文件夹下.

3. 从Release页面下载 预训练模型. 解压到 motion_inbetweening/experiments 文件夹.

4. 安装 PyTorch. 代码在 Python3.8, PyTorch-1.8.2 测试过.

运行 Baseline Benchmark

在 scripts 文件夹下, 运行 python run_baseline_benchmark.py lafan1_context_model

这会给你 Robust Motion In-betweening (Harvey et al., 2020) 论文中展示的相同的 baseline 结果. 如果 LAFAN1 数据集适当地设置, 你将期望看到下面的结果:

trans:  5
zerov_pos: 1.5231, zerov_quat: 0.56, zerov_npss: 0.0053
inter_pos: 0.3729, inter_quat: 0.22, inter_npss: 0.0023
trans: 15
zerov_pos: 3.6946, zerov_quat: 1.10, zerov_npss: 0.0522
inter_pos: 1.2489, inter_quat: 0.62, inter_npss: 0.0391
trans: 30
zerov_pos: 6.6005, zerov_quat: 1.51, zerov_npss: 0.2318
inter_pos: 2.3159, inter_quat: 0.98, inter_npss: 0.2013
trans: 45
zerov_pos: 9.3293, zerov_quat: 1.81, zerov_npss: 0.4918
inter_pos: 3.4471, inter_quat: 1.25, inter_npss: 0.4493

Generate Transition

To use the full method (Detail + Context Transformer) to generate in-betweening, run eval_detail_model.py.

Usage:

usage: eval_detail_model.py [-h] [-s DATASET] [-i INDEX] [-t TRANS] [-d] [-p] det_config ctx_config

Evaluate detail model. No post-processing applied by default.

positional arguments:
  det_config            detail config name
  ctx_config            context config name

optional arguments:
  -h, --help            show this help message and exit
  -s DATASET, --dataset DATASET
                        dataset name (default=benchmark)
  -i INDEX, --index INDEX
                        data index
  -t TRANS, --trans TRANS
                        transition length (default=30)
  -d, --debug           debug mode
  -p, --post_processing
                        apply post-processing

Examples:

1. Get benchmark stats on LAFAN1 dataset with transition=5 frames:

   python eval_detail_model.py lafan1_detail_model lafan1_context_model -t 5

   You are expected to see the same stats shown in our paper:

   trans 5: gpos: 0.1049, gquat: 0.0994, npss: 0.0011

   Try other transition lengths and you should get:

   trans 15: gpos: 0.3943, gquat: 0.2839, npss: 0.0188
   trans 30: gpos: 0.8948, gquat: 0.5446, npss: 0.1124
   trans 45: gpos: 1.6777, gquat: 0.8727, npss: 0.3217

2. Generate 30 transition frames based on the clip with index=100 in LAFAN1 benchmark dataset:

   python eval_detail_model.py lafan1_detail_model lafan1_context_model -t 30 -i 100

   You should get the generated transition and the corresponding ground truth in JSON format under the scripts folder:

   lafan1_detail_model_constraints_benchmark_30_100.json
   lafan1_detail_model_constraints_benchmark_30_100_gt.json

Generate Transition by Context Transformer Only

If you prefer to use only Context Transformer, run eval_context_model.py. Its usage is very similar to eval_detail_model.py. Run python eval_context_model.py -h to see its usage info.

Examples:

1. Get benchmark stats on LAFAN1 dataset with transition=5 frames.

   Context Transformer only, WITHOUT post-processing:

   python eval_context_model.py lafan1_context_model -t 5

   trans 5: gpos: 0.1717, gquat: 0.1325, npss: 0.0015

   Results of other transition lengths:

   trans 15: gpos: 0.4923, gquat: 0.3287, npss: 0.0212
   trans 30: gpos: 1.0663, gquat: 0.5991, npss: 0.1238
   trans 45: gpos: 1.9972, gquat: 0.9170, npss: 0.3369

   Context Transformer only, WITH post-processing:

   python eval_context_model.py lafan1_context_model -t 5 -p

   trans 5: gpos: 0.1288, gquat: 0.1143, npss: 0.0015 (w/ post-processing)

   Results of other transition lengths:

   trans 15: gpos: 0.4623, gquat: 0.3154, npss: 0.0211 (w/ post-processing)
   trans 30: gpos: 1.0354, gquat: 0.5898, npss: 0.1210 (w/ post-processing)
   trans 45: gpos: 1.9439, gquat: 0.9114, npss: 0.3349 (w/ post-processing)

2. Generate 30 transition frames based on the clip with index=100 in LAFAN1 benchmark dataset with post-processing:

   python eval_context_model.py lafan1_context_model -t 30 -i 100 -p

   You should get the predicted transition and the ground truth in JSON format under the scripts folder:

   lafan1_context_model_constraints_benchmark_30_100.json
   lafan1_context_model_constraints_benchmark_30_100_gt.json

Visualize Output Motion in Autodesk Maya

Use the visualize function in motion_inbetween.visualization.maya module to load motions in JSON format:

Training From Scratch

If you want to train the models by yourself, install visdom to visualize training statistics.

pip install visdom

Launch visdom local server before training starts:

$ visdom
Checking for scripts.
It's Alive!

First train the Context Transformer by running train_context_model.py.

usage: train_context_model.py [-h] config

Train context model.

positional arguments:
config      config name

optional arguments:
-h, --help  show this help message and exit

Example:

python train_context_model.py lafan1_context_model

Then train Detail Transformer by running train_detail_model.py.

usage: train_detail_model.py [-h] det_config ctx_config

Train detail model.

positional arguments:
det_config  detail config name
ctx_config  context config name

optional arguments:
-h, --help  show this help message and exit

Example:

python train_detail_model.py lafan1_detail_model lafan1_context_model