目录

关于数据集

创新点

数据处理：

motion可视化代码

---

FineDance：细粒度编舞数据集用于 3D 全身舞蹈生成

关于数据集

目前，最流行的 3D 编舞数据集是 AIST++ [24]，它提供 5 小时的舞蹈，但没有手部动作。AIST++通过多视角视频进行重构。因此，由于重建误差，生成的3D舞蹈动作与真实动作之间不可避免地存在偏差

FineDance 是最大的音乐-舞蹈配对数据集，拥有最多的舞蹈流派

创新点

提出一个 Genre&Coherent aware 检索模块。此外，我们提出了一种新的指标，称为流派匹配分数

motion可视化代码

 

import argparse
import os
from pathlib import Path
import smplx, pickle
import torch
import sysfrom pytorch3d.transforms import rotation_6d_to_matrix, matrix_to_quaternion
from tqdm import tqdm
import glob
import numpy as npsys.path.append(os.getcwd()) 
from dataset.quaternion import ax_to_6v, ax_from_6v
from dataset.preprocess import Normalizer, vectorize_manyimport numpy as np
import matplotlib.pyplot as plt
from mpl_toolkits.mplot3d import Axes3D
from matplotlib.animation import FuncAnimationconnections = [(0, 1),  # 头部到颈部(1, 2), (2, 3), (3, 4),  # 右臂(1, 5), (5, 6), (6, 7),  # 左臂(1, 8),  # 颈部到身体中心(8, 9), (9, 10), (10, 11),  # 右腿(8, 12), (12, 13), (13, 14),  # 左腿(11, 15),  # 右脚到右脚尖(14, 16),  # 左脚到左脚尖
]def motion_feats_extract(inputs_dir, outputs_dir):device = "cuda:0"print("extracting")raw_fps = 30data_fps = 30data_fps <= raw_fpsos.makedirs(outputs_dir,exist_ok=True)# All motion is retargeted to this standard model.motions = sorted(glob.glob(os.path.join(inputs_dir, "*.npy")))for motion in tqdm(motions):name = os.path.splitext(os.path.basename(motion))[0].split(".")[0]print("name is", name)data = np.load(motion, allow_pickle=True)print(data.shape)pos = data[:,:3]   # length, cq = data[:,3:]root_pos = torch.Tensor(pos).to(device) # T, 3length = root_pos.shape[0]local_q_rot6d = torch.Tensor(q).to(device)    # T, 312print("local_q_rot6d", local_q_rot6d.shape)local_q = local_q_rot6d.reshape(length, 52, 6).clone()mat = rotation_6d_to_matrix(local_q)quaternions=matrix_to_quaternion(mat)norms = torch.norm(quaternions[..., 1:], p=2, dim=-1, keepdim=True)half_angles = torch.atan2(norms, quaternions[..., :1])angles = 2 * half_angleseps = 1e-6small_angles = angles.abs() < epssin_half_angles_over_angles = torch.empty_like(angles)sin_half_angles_over_angles[~small_angles] = (torch.sin(half_angles[~small_angles]) / angles[~small_angles])sin_half_angles_over_angles[small_angles] = (0.5 - (angles[small_angles] * angles[small_angles]) / 48)# 这里返回规范化的旋转轴向量rotation_axis_normalized = quaternions[..., 1:] / sin_half_angles_over_angles# norms = torch.norm(quaternions[..., 1:], p=2, dim=-1, keepdim=True)# half_angles = torch.atan2(norms, quaternions[..., :1])# angles = 2 * half_angles# eps = 1e-6# small_angles = angles.abs() < eps# sin_half_angles_over_angles = torch.empty_like(angles)# sin_half_angles_over_angles[~small_angles] = (torch.sin(half_angles[~small_angles]) / angles[~small_angles])# # for x small, sin(x/2) is about x/2 - (x/2)^3/6# # so sin(x/2)/x is about 1/2 - (x*x)/48# sin_half_angles_over_angles[small_angles] = (0.5 - (angles[small_angles] * angles[small_angles]) / 48)# local_q= quaternions[..., 1:] / sin_half_angles_over_angles# local_q = ax_from_6v(local_q).view(length, 156)           # T, 156rotation_axis_normalized=rotation_axis_normalized.cpu().numpy()*10def update_pose(frame, pose,lines):pose=rotation_axis_normalized[frame]# pose[:, 1] = pose[:, 1] + np.sin(frame / 10.0) * 0.02  # 左右摆动# 更新线段以连接关键点for line, (i, j) in zip(lines, connections):line.set_data([pose[i, 0], pose[j, 0]], [pose[i, 1], pose[j, 1]])line.set_3d_properties([pose[i, 2], pose[j, 2]])return lines# 创建动画fig = plt.figure()ax = fig.add_subplot(111, projection='3d')# 设置坐标轴的显示范围ax.set_xlim((0, 10))ax.set_ylim((-6, 6))ax.set_zlim((-5, 20))# pose = np.asarray(init_poses).astype(np.float32)pose = rotation_axis_normalized[0]lines = [ax.plot([pose[s, 0], pose[e, 0]], [pose[s, 1], pose[e, 1]], [pose[s, 2], pose[e, 2]])[0] for s, e in connections]ani = FuncAnimation(fig, update_pose, frames=np.arange(0, rotation_axis_normalized.shape[0]), fargs=(pose,lines), interval=20)plt.show()if __name__ == "__main__":in_dir=r"E:\迅雷下载\data\finedance\motion"out_dir=r"E:\迅雷下载\data\finedance\/motion_fea319"motion_feats_extract(in_dir, out_dir)