egocentric-kitchen-sample / observation_reader.py
dec0y777's picture
Publish three owned egocentric RGB-D kitchen captures under CC BY 4.0
75ccd3f verified
Raw
History Blame Contribute Delete
4.6 kB
"""Standalone reader for Diffraction human egocentric RGB-D samples.
Dependencies: numpy, polars, opencv-python. No pipeline or robot software needed.
"""
from __future__ import annotations
import json
import zipfile
from pathlib import Path
import cv2
import numpy as np
import polars as pl
class ObservationDataset:
def __init__(self, root):
self.root = Path(root)
self.info = json.loads((self.root / "dataset.json").read_text(encoding="utf-8"))
self.episodes = {e["capture_id"]: e for e in self.info["episodes"]}
def table(self, capture_id, name):
e = self.episodes[capture_id]
return pl.read_parquet(self.root / e["signals"][name])
def frame(self, capture_id, frame_index):
"""Return aligned native RGB, metric depth, confidence, K and camera pose.
ARKit pose is device VIO; no ground-truth or robot-action claim is implied.
Depth zero is invalid. Confidence filtering is explicit and reproducible.
"""
e = self.episodes[capture_id]
if not isinstance(frame_index, int) or not 0 <= frame_index < e["frame_count"]:
raise IndexError(frame_index)
mapping = self.table(capture_id, "frame_mapping").filter(pl.col("rgb_frame_index") == frame_index).row(0, named=True)
source_frame_index = mapping["source_frame_index"]
intr = self.table(capture_id, "camera_intrinsics").filter(pl.col("source_frame_index") == source_frame_index).row(0, named=True)
pose = self.table(capture_id, "arkit_poses").filter(pl.col("t_s") == mapping["t_s"]).row(0, named=True)
cap = cv2.VideoCapture(str(self.root / e["video"]))
try:
cap.set(cv2.CAP_PROP_POS_FRAMES, frame_index)
ok, bgr = cap.read()
finally:
cap.release()
if not ok: raise RuntimeError(f"RGB decode failed at frame {frame_index}")
stem = f"{source_frame_index:06d}.png"
with zipfile.ZipFile(self.root / e["sensors"]) as z:
depth_raw = cv2.imdecode(np.frombuffer(z.read("depth/" + stem), np.uint8), cv2.IMREAD_UNCHANGED)
conf = cv2.imdecode(np.frombuffer(z.read("confidence/" + stem), np.uint8), cv2.IMREAD_UNCHANGED)
if depth_raw is None or conf is None or depth_raw.shape != conf.shape:
raise ValueError("Invalid depth/confidence pair")
depth = depth_raw.astype(np.float32) / 1000.
valid = (depth > 0) & (conf >= self.info["min_depth_confidence"])
depth[~valid] = 0
K = np.array([[intr["fx"], 0, intr["cx"]], [0, intr["fy"], intr["cy"]], [0, 0, 1]], np.float64)
return {"rgb": cv2.cvtColor(bgr, cv2.COLOR_BGR2RGB), "depth_m": depth,
"depth_valid": valid, "confidence": conf, "K_rgb": K,
"camera_pose_xyzw": np.array([pose[k] for k in ("tx", "ty", "tz", "qx", "qy", "qz", "qw")]),
"timestamp_s": mapping["t_s"], "sensor_timestamp_s": mapping["sensor_timestamp_s"],
"instruction": e["instruction"], "source_frame_index": source_frame_index, "rgb_frame_index": frame_index}
def annotations_near(self, capture_id, channel, timestamp_s, max_age_s=0.05):
"""Sparse annotations stay absent beyond the caller's explicit age gate."""
df = self.table(capture_id, channel)
if df.is_empty(): return df
distance = (pl.col("t_s") - timestamp_s).abs()
nearby = df.filter(distance <= max_age_s)
if nearby.is_empty(): return nearby
nearest = nearby.select((pl.col("t_s") - timestamp_s).abs().arg_min()).item()
return nearby.filter(pl.col("t_s") == nearby["t_s"][nearest])
def camera_points(frame):
"""Valid depth pixels backprojected in OpenCV camera axes (+x right,+y down,+z forward)."""
depth, K = frame["depth_m"], frame["K_rgb"].copy()
dh, dw = depth.shape; rh, rw = frame["rgb"].shape[:2]
K[0, :] *= dw / rw; K[1, :] *= dh / rh
yy, xx = np.indices(depth.shape)
xyz = np.stack([(xx-K[0, 2])*depth/K[0, 0], (yy-K[1, 2])*depth/K[1, 1], depth], -1)
return xyz[frame["depth_valid"]]
if __name__ == "__main__":
import argparse
p = argparse.ArgumentParser(); p.add_argument("root", type=Path)
args = p.parse_args(); ds = ObservationDataset(args.root)
for key, episode in ds.episodes.items():
frame = ds.frame(key, episode["frame_count"] // 2)
points = camera_points(frame)
print(key, frame["rgb"].shape, frame["depth_m"].shape, len(points), "valid camera points")