File size: 8,044 Bytes
c1a4584 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 | """Smooth per-frame object trajectory from an optimized layout JSON.
Filters out OOD frames (jump-and-jump-back outliers) via interpolation,
then applies Savitzky-Golay filtering to the translation and rotation
(quaternion, wxyz) fields. Only those two fields are modified; every other
field in the layout is passed through unchanged.
By default this operates on the camera-frame pose produced by
`convert_layout_to_camera_frame.py` + `optimize_translation_scale.py`
(`translation_camera_frame` / `quat_wxyz_camera_frame`), which is what the
downstream visualizer/retargeter consumes. Use --trans-key / --quat-key to
smooth the pose-frame fields (`translation` / `new_quat`) instead.
"""
import argparse
import json
import copy
import os
import numpy as np
from pathlib import Path
from scipy.signal import savgol_filter
from scipy.spatial.transform import Rotation, Slerp
# --- Defaults ---
WINDOW_SIZE = 7 # must be odd
POLY_ORDER = 3
# OOD detection: frame is outlier if both its incoming and outgoing velocity
# exceed this many times the median velocity
OOD_MULTIPLIER = 3.0
# Field keys (camera-frame pose is the pipeline's final consumed pose)
DEFAULT_TRANS_KEY = "translation_camera_frame"
DEFAULT_QUAT_KEY = "quat_wxyz_camera_frame"
def detect_ood_frames(values: np.ndarray, threshold_mult: float) -> list[int]:
"""Detect frames that jump out and back (OOD spikes).
A frame i is OOD if the change from i-1->i AND i->i+1 are both
above threshold_mult * median change.
"""
deltas = np.linalg.norm(np.diff(values, axis=0), axis=1)
median_delta = np.median(deltas)
threshold = max(threshold_mult * median_delta, 1e-6)
ood = []
for i in range(len(deltas) - 1):
if deltas[i] > threshold and deltas[i + 1] > threshold:
ood.append(i + 1) # the frame that jumped
return ood
def interpolate_ood_translation(trans: np.ndarray, ood_frames: list[int]) -> np.ndarray:
"""Replace OOD frames with linearly interpolated values."""
result = trans.copy()
for i in ood_frames:
# Find nearest non-OOD neighbors
lo = i - 1
while lo in ood_frames and lo > 0:
lo -= 1
hi = i + 1
while hi in ood_frames and hi < len(trans) - 1:
hi += 1
if lo == hi:
continue
t = (i - lo) / (hi - lo)
result[i] = (1 - t) * result[lo] + t * result[hi]
return result
def interpolate_ood_rotation(quats_xyzw: np.ndarray, ood_frames: list[int]) -> np.ndarray:
"""Replace OOD frames with SLERP-interpolated rotations."""
result = quats_xyzw.copy()
ood_set = set(ood_frames)
for i in ood_frames:
lo = i - 1
while lo in ood_set and lo > 0:
lo -= 1
hi = i + 1
while hi in ood_set and hi < len(quats_xyzw) - 1:
hi += 1
if lo == hi:
continue
rots = Rotation.from_quat([result[lo], result[hi]])
slerp = Slerp([0, 1], rots)
t = (i - lo) / (hi - lo)
result[i] = slerp([t]).as_quat()[0]
return result
def smooth_trajectory(input_path: Path, output_path: Path, trans_key: str, quat_key: str):
with open(input_path) as f:
data = json.load(f)
def get_frame_idx(o):
return o.get("frame_index", o.get("frame_idx"))
# Separate dynamic (has a frame index) and static objects
dynamic = [o for o in data["objects"] if get_frame_idx(o) is not None]
static = [o for o in data["objects"] if get_frame_idx(o) is None]
# Sort by frame index
dynamic.sort(key=get_frame_idx)
n = len(dynamic)
print(f"Dynamic frames: {n}, Static objects: {len(static)}")
print(f"Smoothing fields: translation='{trans_key}', quaternion='{quat_key}'")
# Extract arrays
trans = np.array([o["local_to_scene"][trans_key] for o in dynamic]) # (N, 3)
quats_wxyz = np.array([o["local_to_scene"][quat_key] for o in dynamic]) # (N, 4) wxyz
# Convert quat wxyz -> scipy xyzw
quats_xyzw = quats_wxyz[:, [1, 2, 3, 0]]
# Ensure quaternion continuity
for i in range(1, n):
if np.dot(quats_xyzw[i], quats_xyzw[i - 1]) < 0:
quats_xyzw[i] = -quats_xyzw[i]
# --- Detect OOD frames ---
rotvec = Rotation.from_quat(quats_xyzw).as_rotvec()
ood_trans = detect_ood_frames(trans, OOD_MULTIPLIER)
ood_rot = detect_ood_frames(rotvec, OOD_MULTIPLIER)
ood_all = sorted(set(ood_trans + ood_rot))
print(f"OOD translation frames: {ood_trans}")
print(f"OOD rotation frames: {ood_rot}")
print(f"Combined OOD frames: {ood_all}")
# --- Replace OOD frames with interpolated values ---
trans_clean = interpolate_ood_translation(trans, ood_all)
quats_clean = interpolate_ood_rotation(quats_xyzw, ood_all)
# --- Smooth translation ---
trans_smooth = np.zeros_like(trans_clean)
for i in range(3):
trans_smooth[:, i] = savgol_filter(trans_clean[:, i], WINDOW_SIZE, POLY_ORDER)
# --- Smooth quaternion directly in R^4 and renormalize ---
# Works because sign-continuity is enforced above; no angle-wrap issues.
quats_smooth = np.zeros_like(quats_clean)
for i in range(4):
quats_smooth[:, i] = savgol_filter(quats_clean[:, i], WINDOW_SIZE, POLY_ORDER)
quats_smooth /= np.linalg.norm(quats_smooth, axis=1, keepdims=True)
quats_xyzw_smooth = quats_smooth
quats_wxyz_smooth = quats_xyzw_smooth[:, [3, 0, 1, 2]]
# --- Build output (only modify the translation and quaternion fields) ---
output = copy.deepcopy(data)
smoothed_dynamic = copy.deepcopy(dynamic)
for i, obj in enumerate(smoothed_dynamic):
pose = obj["local_to_scene"]
pose[trans_key] = trans_smooth[i].tolist()
pose[quat_key] = quats_wxyz_smooth[i].tolist()
output["objects"] = smoothed_dynamic + copy.deepcopy(static)
os.makedirs(os.path.dirname(os.path.abspath(output_path)), exist_ok=True)
with open(output_path, "w") as f:
json.dump(output, f, indent=2)
print(f"\nWrote smoothed trajectory to {output_path}")
# --- Summary statistics ---
trans_diff = np.abs(trans_smooth - trans)
print(f"\nTranslation smoothing (mean abs change per axis):")
print(f" x: {trans_diff[:, 0].mean():.6f} y: {trans_diff[:, 1].mean():.6f} z: {trans_diff[:, 2].mean():.6f}")
print(f" max change: {trans_diff.max():.6f}")
angle_diff = [
(Rotation.from_quat(quats_xyzw[i]).inv() * Rotation.from_quat(quats_xyzw_smooth[i])).magnitude()
for i in range(n)
]
angle_diff = np.array(angle_diff)
print(f"\nQuaternion smoothing (angle change in radians):")
print(f" mean: {angle_diff.mean():.6f} max: {angle_diff.max():.6f}")
print(f"\nSample translations (frame: original -> smoothed):")
for idx in [0, n // 4, n // 2, 3 * n // 4, n - 1]:
print(f" frame {get_frame_idx(dynamic[idx])}: "
f"{trans[idx].tolist()} -> {trans_smooth[idx].tolist()}")
def parse_args():
p = argparse.ArgumentParser(description="Smooth per-frame object trajectory")
p.add_argument("--input", type=str, required=True, help="Input layout JSON path")
p.add_argument("--output", type=str, default=None,
help="Output smoothed layout JSON path (default: <input>_smooth.json)")
p.add_argument("--trans-key", type=str, default=DEFAULT_TRANS_KEY,
help="local_to_scene translation field to smooth "
f"(default: {DEFAULT_TRANS_KEY}; use 'translation' for pose frame)")
p.add_argument("--quat-key", type=str, default=DEFAULT_QUAT_KEY,
help="local_to_scene quaternion (wxyz) field to smooth "
f"(default: {DEFAULT_QUAT_KEY}; use 'new_quat' for pose frame)")
return p.parse_args()
if __name__ == "__main__":
args = parse_args()
output = args.output
if output is None:
stem, ext = os.path.splitext(args.input)
output = f"{stem}_smooth{ext}"
smooth_trajectory(args.input, output, args.trans_key, args.quat_key)
|