Could you elaborate, please, are you using Motion-X format for presenting 668 features in .npy files? If no, can u provide the script how you got the features from SMPL-X parameters. Thanks in advance!
", "messages": [ { "id": "3425816", "postDate": "03/21/2026 10:32:20", "content": "Could you elaborate, please, are you using Motion-X format for presenting 668 features in .npy files? If no, can u provide the script how you got the features from SMPL-X parameters. Thanks in advance!
", "rawMarkdown": "Could you elaborate, please, are you using Motion-X format for presenting 668 features in .npy files? If no, can u provide the script how you got the features from SMPL-X parameters. Thanks in advance!", "votes": null }, { "id": "3426779", "postDate": "03/23/2026 07:51:55", "content": "We did not use Motion-X format. We first converted the SMPL-X parameters to BVH using the same logic as this smpl2bvh script, extracting rotations and root translation from the .npz/.pkl files and writing them into a BVH skeleton with 55 joints. We then stitched per-word BVH clips into per-sentence BVH files with smooth blending at transitions, and ran a HumanML3D-style feature extraction pipeline:
\nForward kinematics → global joint positions\nFloor alignment (character placed on ground plane)\nForward-direction extraction + Gaussian smoothing\nRoot rotation canonicalization (all poses face Z+)\nFeature concatenation:\n4D root - angular velocity (1D) + X/Z linear velocity (2D) + height (1D)\n330D rotations - all 55 joints in 6D continuous rotation representation (Zhou et al. 2019)\n165D positions - root-invariant joint positions (55 × 3)\n165D velocities - frame-to-frame joint velocities (55 × 3)\n4D foot contacts - binary contact labels for left/right ankle and foot\nTotal: 4 + 330 + 165 + 165 + 4 = 668
", "rawMarkdown": "We did not use Motion-X format. We first converted the SMPL-X parameters to BVH using the same logic as this smpl2bvh script, extracting rotations and root translation from the .npz/.pkl files and writing them into a BVH skeleton with 55 joints. We then stitched per-word BVH clips into per-sentence BVH files with smooth blending at transitions, and ran a HumanML3D-style feature extraction pipeline:\n\nForward kinematics → global joint positions\nFloor alignment (character placed on ground plane)\nForward-direction extraction + Gaussian smoothing\nRoot rotation canonicalization (all poses face Z+)\nFeature concatenation:\n4D root - angular velocity (1D) + X/Z linear velocity (2D) + height (1D)\n330D rotations - all 55 joints in 6D continuous rotation representation (Zhou et al. 2019)\n165D positions - root-invariant joint positions (55 × 3)\n165D velocities - frame-to-frame joint velocities (55 × 3)\n4D foot contacts - binary contact labels for left/right ankle and foot\nTotal: 4 + 330 + 165 + 165 + 4 = 668", "votes": null }, { "id": "3430293", "postDate": "03/28/2026 07:36:52", "content": "Thank you for providing this information!
", "rawMarkdown": "Thank you for providing this information!", "votes": null } ], "comments": [ { "id": 3426779, "author_name": "antonygithinji", "author_url": "", "post_date": "03/23/2026 07:51:55", "content": "We did not use Motion-X format. We first converted the SMPL-X parameters to BVH using the same logic as this smpl2bvh script, extracting rotations and root translation from the .npz/.pkl files and writing them into a BVH skeleton with 55 joints. We then stitched per-word BVH clips into per-sentence BVH files with smooth blending at transitions, and ran a HumanML3D-style feature extraction pipeline:
\nForward kinematics → global joint positions\nFloor alignment (character placed on ground plane)\nForward-direction extraction + Gaussian smoothing\nRoot rotation canonicalization (all poses face Z+)\nFeature concatenation:\n4D root - angular velocity (1D) + X/Z linear velocity (2D) + height (1D)\n330D rotations - all 55 joints in 6D continuous rotation representation (Zhou et al. 2019)\n165D positions - root-invariant joint positions (55 × 3)\n165D velocities - frame-to-frame joint velocities (55 × 3)\n4D foot contacts - binary contact labels for left/right ankle and foot\nTotal: 4 + 330 + 165 + 165 + 4 = 668
", "votes": null, "replies": [ { "id": 3430293, "author_name": "maksimrazantsau", "author_url": "", "post_date": "03/28/2026 07:36:52", "content": "Thank you for providing this information!
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "3425816": "Could you elaborate, please, are you using Motion-X format for presenting 668 features in .npy files? If no, can u provide the script how you got the features from SMPL-X parameters. Thanks in advance!", "3426779": "We did not use Motion-X format. We first converted the SMPL-X parameters to BVH using the same logic as this smpl2bvh script, extracting rotations and root translation from the .npz/.pkl files and writing them into a BVH skeleton with 55 joints. We then stitched per-word BVH clips into per-sentence BVH files with smooth blending at transitions, and ran a HumanML3D-style feature extraction pipeline:\n\nForward kinematics → global joint positions\nFloor alignment (character placed on ground plane)\nForward-direction extraction + Gaussian smoothing\nRoot rotation canonicalization (all poses face Z+)\nFeature concatenation:\n4D root - angular velocity (1D) + X/Z linear velocity (2D) + height (1D)\n330D rotations - all 55 joints in 6D continuous rotation representation (Zhou et al. 2019)\n165D positions - root-invariant joint positions (55 × 3)\n165D velocities - frame-to-frame joint velocities (55 × 3)\n4D foot contacts - binary contact labels for left/right ankle and foot\nTotal: 4 + 330 + 165 + 165 + 4 = 668", "3430293": "Thank you for providing this information!" }, "source": "meta" }