movement-strict-164: High-Quality Filtered Pose Dataset

164,390 clips from Kinetics-700 that pass both programmatic continuity checks and a 235B-parameter Vision-Language Model judge that evaluated the rendered skeleton overlay against the action label. Roughly 60% of clips have been re-tracked through a dedicated multi-frame YOLO + Qwen oracle + sticky IoU tracker pipeline before judgment, replacing the original tracking with a cleaner result.

This is a filtered subset of maxsegan/movenet-332 intended as the highest-quality tier for training Vision-Language-Action models for humanoid robots. The companion broader release with the same retracking but a more permissive quality bar is at maxsegan/movement-287.

What "strict" means

Each clip's tracking was re-judged with Qwen3-VL-235B-A22B-Thinking-FP8 looking at six rendered overlay frames per clip. A clip is strict if all of these hold:

• v2_prog_good — passes continuity / motion / size checks (no hard cuts, no IoU breaks, no tracking jumps, peak joint speed ≥ 3 m/s, bbox area ≥ 8% of frame, etc.)
• v2_vlm_tracking_good — VLM confirms the same person is cleanly tracked throughout
• v2_vlm_motion_matches — VLM confirms the tracked person is performing the labeled action

This drops slow legitimate motion (texting, archaeological excavation) along with tracking artifacts, in exchange for the cleanest possible velocity profiles and identity continuity. Use the broader movement-287 set if you want to keep slower actions.

Re-tracking pipeline

For each clip flagged by either signal in a first pass, the original tracking was replaced via:

1. Multi-frame YOLOv8x candidate pool over 5 sampled frames
2. Qwen3-VL oracle picks the subject on the busiest anchor frame
3. Sticky IoU tracker (threshold 0.5, size-ratio guard 2.5×, 3-miss tolerance)
4. Trim to longest valid run containing the anchor; re-detect hard cuts on the trimmed range
5. Interpolate residual interior NaN bbox gaps
6. ViTPose-plus-large 2D + MotionAGFormer-B 3D lifting on the new tracking

The retracked clips ship with shorter num_pose_frames than the originals when the trimmer cut off bad segments — frame ranges 35–250 are typical.

Schema

Column	Type	Description
clip_id	string	NPZ filename stem (e.g., a4UOhPiV4QE_000675_000685)
action_class	string	Kinetics-700 action class name
youtube_id	string	YouTube video ID
time_start, time_end	string	Start/end timestamp from the clip ID
split	string	Original train/val split (preserved from movenet-332)
instruction	string	VLM-generated imperative caption
fps	float32	Pose-data FPS (20.0 for retracked clips, original for source=original)
num_pose_frames	int32	Number of pose frames after retrack/trim
video_width, video_height	int32	Original video dimensions
pose3d	bytes	zlib-compressed float32 [F, 17, 3] H36M format
keypoints2d	bytes	zlib-compressed float32 [F, 17, 2]
scores2d	bytes	zlib-compressed float32 [F, 17] joint confidences
bboxes	bytes	zlib-compressed float32 [F, 4] [x1, y1, x2, y2]
joint_angles	bytes	empty for retracked clips (recompute from pose3d); preserved zlib-compressed float32 [F, 22] for source=original
frame_indices	bytes	zlib-compressed int32 [F] mapping pose frame → video frame
tracking_confidence	bytes	zlib-compressed float32 [F] per-frame mean joint confidence
has_hard_cuts	bool	Whether hard cuts were detected in the (possibly retracked) sequence
quality	float32	Original quality score; superseded by v2_* columns
v2_source	string	"retrack" or "original"
v2_prog_good	bool	Programmatic check verdict
v2_vlm_tracking_good	bool	235B VLM tracking-consistency verdict
v2_vlm_motion_matches	bool	235B VLM motion-matches-label verdict
v2_vlm_good	bool	vlm_tracking_good AND vlm_motion_matches
v2_strict	bool	prog_good AND vlm_good — always True in this release
v2_prog_issues	list[string]	Names of failing programmatic checks (empty in this release)

Array columns are zlib-compressed raw numpy bytes. Variable-length across clips.

Loading

from datasets import load_dataset
import numpy as np
import zlib

ds = load_dataset("maxsegan/movement-strict-164", split="train")
row = ds[0]
F = row["num_pose_frames"]
pose3d = np.frombuffer(zlib.decompress(row["pose3d"]),
                       dtype=np.float32).reshape(F, 17, 3)
print(row["clip_id"], row["action_class"], row["v2_source"], pose3d.shape)

Source breakdown (164,390 clips)

Source	Count	Description
original	107,914	Phase-1 strict clips that already passed both signals on the original parquet — tracking unchanged from movenet-332
retrack	56,476	Phase-2 clips that became strict only after re-tracking

Composition stats vs movenet-332

	movenet-332 (raw)	movement-strict-164
Clips	325,036	164,390
Mean clip duration	~24 s	~7 s (after trim-to-valid-run)
Action classes	700	700
Min clips/class	12	varies

Skeleton & joint angle conventions

H36M 17-joint skeleton, same as movenet-332. See maxsegan/movenet-332 for the joint hierarchy and 22-DoF layout.

To recompute joint angles from pose3d (the source=retrack rows have empty joint_angles), use the helper at training/kinetics_dataset.py::pose3d_to_joint_angles in the GitHub repo.

Pipeline code

The full re-tracking + judging + filtering pipeline lives at https://github.com/maxsegan/movement (scripts/retrack_clips.py, scripts/scale_judge.py, scripts/scale_judge_npz.py, scripts/compose_final_dataset.py, scripts/package_v2.py).

License

Pose annotations and metadata: CC-BY-4.0. Underlying Kinetics-700 videos are subject to their own license; this dataset does not redistribute video content.