hkust-aaron's picture
Add camera motion, relative depth, task graph, scene graphs, action tokens and robot-learning exports; data spec
f20e058 verified
Raw History Blame Contribute Delete
3.66 kB
"""Relative monocular depth for every frame (Depth Anything V2 Small), sampled at tool tips (CUDA machine).
Usage: python 16_depth.py <video> <model_dir> <tool_tips.csv> <dense_masks_dir> <out_dir> <clips e.g. 5,40,43>
Depth Anything V2 predicts relative inverse depth (larger = closer) up to an
unknown scale and shift per frame; it is not metric. To make values comparable
over time each frame is normalised by the median of its background pixels
(instrument masks excluded): rel_inv_depth = prediction / background median.
> 1 closer to the lens than the typical tissue, < 1 farther away.
Per instrument-frame the tip value is the median in a 7x7 window at the tip and
the shaft value the median over the mask.
For the listed clips the full maps are also saved every 5th frame as 16-bit PNG
(360x288, value = rel_inv_depth x 10000).
Writes tool_depth.csv and depth_maps/clip_XX/<frame>.png.
"""
import csv
import sys
from pathlib import Path
import cv2
import numpy as np
import torch
from transformers import AutoImageProcessor, AutoModelForDepthEstimation
video, model_dir, tips_csv, masks_dir, out = sys.argv[1], sys.argv[2], sys.argv[3], Path(sys.argv[4]), Path(sys.argv[5])
save_clips = {int(x) for x in sys.argv[6].split(",")}
out.mkdir(parents=True, exist_ok=True)
proc = AutoImageProcessor.from_pretrained(model_dir)
model = AutoModelForDepthEstimation.from_pretrained(model_dir).cuda().eval().half()
tips = {}
with open(tips_csv) as fh:
for r in csv.DictReader(fh):
tips.setdefault(int(r["frame"]), []).append(r)
cap = cv2.VideoCapture(video)
rows, f, BATCH = [], 0, 16
buf = []
def flush(buf):
imgs = [b[1][..., ::-1] for b in buf]
inp = proc(images=imgs, return_tensors="pt").to("cuda")
with torch.no_grad():
pred = model(pixel_values=inp["pixel_values"].half()).predicted_depth.float()
pred = torch.nn.functional.interpolate(pred[:, None], size=(576, 720), mode="bicubic", align_corners=False)[:, 0]
for (fi, _), d in zip(buf, pred.cpu().numpy()):
inst = cv2.imread(str(masks_dir / f"clip_{fi // 250:02d}" / f"{fi:05d}.png"), 0)
bgm = np.median(d[inst == 0]) if (inst == 0).any() else np.median(d)
rel = d / max(bgm, 1e-6)
for r in tips.get(fi, []):
tid = int(r["track_id"])
m = inst == tid
row = {"frame": fi, "clip_id": r["clip_id"], "track_id": tid, "category": r["category"],
"shaft_rel_inv_depth": round(float(np.median(rel[m])), 4) if m.any() else "",
"tip_rel_inv_depth": ""}
if r["tip_x"]:
x, y = int(round(float(r["tip_x"]))), int(round(float(r["tip_y"])))
win = rel[max(0, y - 3):y + 4, max(0, x - 3):x + 4]
row["tip_rel_inv_depth"] = round(float(np.median(win)), 4)
rows.append(row)
c = fi // 250
if c in save_clips and fi % 5 == 0:
(out / "depth_maps" / f"clip_{c:02d}").mkdir(parents=True, exist_ok=True)
small = cv2.resize(rel, (360, 288), interpolation=cv2.INTER_AREA)
cv2.imwrite(str(out / "depth_maps" / f"clip_{c:02d}" / f"{fi:05d}.png"),
np.clip(small * 10000, 0, 65535).astype(np.uint16))
while True:
ok, img = cap.read()
if not ok:
break
buf.append((f, img))
f += 1
if len(buf) == BATCH:
flush(buf)
buf = []
if f % 2000 == 0:
print("frames", f, flush=True)
if buf:
flush(buf)
with open(out / "tool_depth.csv", "w", newline="") as fh:
w = csv.DictWriter(fh, list(rows[0]))
w.writeheader()
w.writerows(rows)
print("done frames", f, "rows", len(rows))