For anyone would like visualize pkl animation without blender, here is a straight-forward way!

[luohao123]

By using pip install nosmpl

I found this script can visualize HybrIK pkl file directly:

import numpy as np
from nosmpl.vis.vis_o3d import vis_mesh_o3d, Open3DVisualizer
from nosmpl.utils import rot_mat_to_euler, rotmat_to_rotvec, quat2mat, quat_to_rotvec
import sys
import pickle
from nosmpl.smpl_onnx import SMPLOnnxRuntime

def vis_pkl():    
    smpl = SMPLOnnxRuntime()

    data_f = sys.argv[1]
    data = pickle.load(open(data_f, "rb"))
    ks = list(data.keys())
    print(ks)

    o3d_vis = Open3DVisualizer(fps=60, save_img_folder="results/", enable_axis=False)

    frame_len = len(data[ks[0]])
    print("frames: ", frame_len)
    for fi in range(frame_len):
        pose = data['pred_thetas'][fi]
        pose = np.array(pose).reshape(-1, 3, 3)
        print(pose.shape)

        trans = data["transl"][fi]

        pose_rotvec = [rotmat_to_rotvec(p) for p in pose]
        pose_rotvec = np.array(pose_rotvec).reshape(-1, 3)
        print(pose_rotvec.shape)

        # global_orient = pose_rotvec[:, :3].astype(np.float32)
        global_orient = np.array([[0, 0, 0]], dtype=np.float32).reshape([1, 1, 3])
        # global_orient = [[i[0], -i[1], i[2]] for i in global_orient]
        # global_orient = np.array(global_orient).astype(np.float32)

        pose_rotvec_body = np.delete(pose_rotvec, [-1], axis=0).reshape(1, 23, 3)
        body = pose_rotvec_body.astype(np.float32)
        # lhand = np.zeros([1, 45]).astype(np.float32)
        # rhand = np.zeros([1, 45]).astype(np.float32)

        outputs = smpl.forward(body, global_orient)

        vertices, joints, faces = outputs
        vertices = vertices[0].squeeze()
        joints = joints[0].squeeze()

        faces = faces.astype(np.int32)
        # vis_mesh_o3d(vertices, faces)
        # vertices += trans
        # trans = [trans[1], trans[0], trans[2]]
        trans = [trans[0], trans[1], 0]
        print(trans)
        o3d_vis.update(vertices, faces, trans, R_along_axis=[np.pi, 0, 0], waitKey=1)
    o3d_vis.release()

if __name__ == "__main__":
    vis_pkl()

Hope it helps!

[chris-hndz]

How should this code be modified to display the SMPL-X pkl animation?

Thank you in advance! 😄

[Brian417-cup]

By using pip install nosmpl

I found this script can visualize HybrIK pkl file directly:

import numpy as np
from nosmpl.vis.vis_o3d import vis_mesh_o3d, Open3DVisualizer
from nosmpl.utils import rot_mat_to_euler, rotmat_to_rotvec, quat2mat, quat_to_rotvec
import sys
import pickle
from nosmpl.smpl_onnx import SMPLOnnxRuntime

def vis_pkl():    
    smpl = SMPLOnnxRuntime()

    data_f = sys.argv[1]
    data = pickle.load(open(data_f, "rb"))
    ks = list(data.keys())
    print(ks)

    o3d_vis = Open3DVisualizer(fps=60, save_img_folder="results/", enable_axis=False)

    frame_len = len(data[ks[0]])
    print("frames: ", frame_len)
    for fi in range(frame_len):
        pose = data['pred_thetas'][fi]
        pose = np.array(pose).reshape(-1, 3, 3)
        print(pose.shape)

        trans = data["transl"][fi]

        pose_rotvec = [rotmat_to_rotvec(p) for p in pose]
        pose_rotvec = np.array(pose_rotvec).reshape(-1, 3)
        print(pose_rotvec.shape)

        # global_orient = pose_rotvec[:, :3].astype(np.float32)
        global_orient = np.array([[0, 0, 0]], dtype=np.float32).reshape([1, 1, 3])
        # global_orient = [[i[0], -i[1], i[2]] for i in global_orient]
        # global_orient = np.array(global_orient).astype(np.float32)

        pose_rotvec_body = np.delete(pose_rotvec, [-1], axis=0).reshape(1, 23, 3)
        body = pose_rotvec_body.astype(np.float32)
        # lhand = np.zeros([1, 45]).astype(np.float32)
        # rhand = np.zeros([1, 45]).astype(np.float32)

        outputs = smpl.forward(body, global_orient)

        vertices, joints, faces = outputs
        vertices = vertices[0].squeeze()
        joints = joints[0].squeeze()

        faces = faces.astype(np.int32)
        # vis_mesh_o3d(vertices, faces)
        # vertices += trans
        # trans = [trans[1], trans[0], trans[2]]
        trans = [trans[0], trans[1], 0]
        print(trans)
        o3d_vis.update(vertices, faces, trans, R_along_axis=[np.pi, 0, 0], waitKey=1)
    o3d_vis.release()

if __name__ == "__main__":
    vis_pkl()

Hope it helps!

Hey, I discovered you directly use root translation in camera space with (x,y,0). Actually, the author has provided a relative code transfer point from camera space into pixel space with corresponding focal length, maybe using the transfer will have a better visualization!!

[jiajiajibamaoss]

(hybrik3d) mocap3d@mocap3D:~/HybrIK$ python scripts/jjj.py Traceback (most recent call last): File "/home/mocap3d/HybrIK/scripts/jjj.py", line 59, in vis_pkl() File "/home/mocap3d/HybrIK/scripts/jjj.py", line 10, in vis_pkl smpl = SMPLOnnxRuntime() File "/home/mocap3d/anaconda3/envs/hybrik3d/lib/python3.9/site-packages/nosmpl/smpl_onnx.py", line 24, in init download('https://github.com/jinfagang/nosmpl/releases/download/v1.1/smpl.onnx', NameError: name 'download' is not defined