Whyx-PROmpTea / src /pose_tagger.py
ArtShumov's picture
feat(tagger): Qwen-VL NL captions, pose extraction (YOLOv8-n), slide-panel model selector
fa6338f
Raw
History Blame
8.34 kB
"""Pose estimation from an image via DWPose/YOLO11n ONNX Runtime.
Runs a lightweight YOLO11n-pose ONNX model for person detection + 17-keypoint
pose estimation. The 17 COCO keypoints are turned into a natural-language
pose tag string that can be appended to a prompt (e.g. ``standing, arms up``).
"""
from __future__ import annotations
import json
import os
from typing import Optional
import numpy as np
from PIL import Image, ImageOps
try:
import onnxruntime as ort
except Exception: # pragma: no cover
ort = None
# ---------------------------------------------------------------------------
# Keypoint analysis
# ---------------------------------------------------------------------------
import os
_DEBUG_POSE = os.environ.get("WHYX_DEBUG_POSE", "0").strip().lower() in ("1", "true", "yes")
_COCO_KP = [
"nose", "left eye", "right eye", "left ear", "right ear", "left shoulder",
"right shoulder", "left elbow", "right elbow", "left wrist", "right wrist",
"left hip", "right hip", "left knee", "right knee", "left ankle",
"right ankle",
]
# YOLOv8-pose ONNX model keypoint order is COCO 17. Each detection row is
# [x, y, w, h, conf, kp0_x, kp0_y, kp0_conf, ..., kp16_x, kp16_y, kp16_conf].
_KPT_START = 5 # detection feature index where keypoints begin
_KPT_STRIDE = 3 # (x, y, conf) per keypoint
def _yolo_to_keypoints(out: np.ndarray, conf_thresh: float = 0.25, iou_thresh: float = 0.5):
"""out shape: (1, N+4, num_detections) — 84 for det-only, 51+5 for pose."""
# (1, C, D) -> (D, C) so each row is a candidate detection.
preds = out[0].T
conf = preds[:, 4]
keep = conf > conf_thresh
preds = preds[keep]
if not len(preds):
return []
# NMS via greedy diagonal covariance check (good enough at pose level).
def _nms(rows, thresh):
order = conf[keep].argsort()[::-1]
keep_idx = []
suppressed = set()
centers = rows[:, :2]
wh = rows[:, 2:4]
areas = wh[:, 0] * wh[:, 1]
for i in order:
if i in suppressed:
continue
keep_idx.append(i)
cx1, cy1 = centers[i]
w1, h1 = wh[i]
for j in order:
if j == i or j in suppressed:
continue
cx2, cy2 = centers[j]
w2, h2 = wh[j]
# Intersection over smaller area (proxy for pose NMS).
dx = max(0, min(cx1 + w1/2, cx2 + w2/2) - max(cx1 - w1/2, cx2 - w2/2))
dy = max(0, min(cy1 + h1/2, cy2 + h2/2) - max(cy1 - h1/2, cy2 - h2/2))
inter = dx * dy
smaller = min(areas[i], areas[j])
if smaller > 0 and inter / smaller > thresh:
suppressed.add(j)
return rows[keep_idx]
preds = _nms(preds, iou_thresh)
kpts = []
for row in preds:
kp = row[_KPT_START:].reshape(-1, _KPT_STRIDE)
kpts.append(kp)
return kpts
def _keypoints_to_pose_tags(kpts: list[np.ndarray]) -> list[str]:
"""Heuristic mapping of COCO-17 coordinates to coarse pose tags.
Uses shoulder-width normalisation so tags stay pose-invariant (scale blur).
`kp[:, 0]` = x, `kp[:, 1]` = y, `kp[:, 2]` = confidence per keypoint.
"""
tags: list[str] = []
if not kpts:
return tags
for kp in kpts: # kp: (17, 3), COCO order
# Shoulder width — pose-invariant scale reference.
ls, rs = kp[5], kp[6]
sw = max(abs(ls[0] - rs[0]), 1e-4)
lh, rh = kp[11], kp[12]
th = max(abs(lh[0] - rh[0]), 1e-4)
lw, rw = kp[9], kp[10]
lk, rk = kp[13], kp[14]
la, ra = kp[15], kp[16]
# ---- arms ----
for side, w, s in (("left", lw, ls), ("right", rw, rs)):
if w[2] < 0.3 or s[2] < 0.3:
continue
# arm up: wrist significantly above shoulder
if w[1] < s[1] - 0.15 * sw:
tags.append(f"{side} arm up")
# arm extended sideways: wrist is far from shoulder laterally
elif abs(w[0] - s[0]) > 0.4 * sw:
tags.append(f"{side} arm out")
# ---- posture: torso vs limb geometry ----
if lh[2] > 0.3 and lk[2] > 0.3:
# sitting: knee is raised toward hip level
if lk[1] < lh[1] - 0.05 * th:
tags.append("sitting")
else:
tags.append("standing")
# lying: torso is more horizontal than vertical
if ls[2] > 0.3 and lh[2] > 0.3:
tv = abs(ls[1] - lh[1])
th_ = abs(ls[0] - lh[0])
if th_ > tv * 1.5:
tags.append("lying")
# walking/running: legs scissor horizontally
if la[2] > 0.3 and ra[2] > 0.3:
stride = abs(la[0] - ra[0])
if stride > 0.8 * sw:
tags.append("walking" if stride < 2.0 * sw else "running")
# kneeling: both knees below hips significantly, sitting
if lk[2] > 0.3 and rk[2] > 0.3 and lh[2] > 0.3:
if lk[1] > lh[1] + 0.4 * th and rk[1] > lh[1] + 0.4 * th:
tags.append("kneeling")
# torso lean: shoulder-to-hip vector strongly non-vertical
if ls[2] > 0.3 and lh[2] > 0.3:
lean = abs(ls[0] - lh[0]) / max(abs(ls[1] - lh[1]), 1e-4)
if lean > 0.35:
tags.append("leaning")
# Deduplicate — for multi-person detections keep only the majority vote.
from collections import Counter
counts = Counter(tags)
return [t for t, n in sorted(counts.items(), key=lambda p: (-p[1], p[0]))]
# ---------------------------------------------------------------------------
# ONNX pose runner
# ---------------------------------------------------------------------------
def _providers() -> list[str]:
avail = ort.get_available_providers()
return ["CUDAExecutionProvider", "CPUExecutionProvider"] if "CUDAExecutionProvider" in avail else ["CPUExecutionProvider"]
class PoseEstimator:
"""YOLO11n-pose wrapper with lazy model download."""
def __init__(self, repo_id: str = "Xenova/yolov8n-pose", filename: str = "onnx/model.onnx"):
self._repo = repo_id
self._filename = filename
self._session: Optional["ort.InferenceSession"] = None
self._input_shape: tuple[int, int] = (640, 640)
self._loaded = False
def ensure_loaded(self) -> bool:
if self._loaded:
return True
if ort is None:
return False
from huggingface_hub import hf_hub_download
try:
model_path = hf_hub_download(repo_id=self._repo, filename=self._filename)
except Exception:
return False
sess = ort.InferenceSession(model_path, providers=_providers())
inp = sess.get_inputs()[0]
shape = inp.shape
if len(shape) == 4:
self._input_shape = (int(shape[2]), int(shape[3]))
self._session = sess
self._loaded = True
return True
def estimate(self, image) -> dict:
"""Return {pose_tags, pose_score, people_count, raw_keypoints_count}."""
if not self.ensure_loaded():
return {"pose_tags": [], "people_count": 0}
pil = image if isinstance(image, Image.Image) else Image.fromarray(np.asarray(image))
pil = ImageOps.exif_transpose(ImageOps.fit(pil, self._input_shape, Image.LANCZOS))
arr = np.asarray(pil, dtype=np.float32) / 255.0
arr = arr.transpose(2, 0, 1)[None, ...] # NCHW
inp_name = self._session.get_inputs()[0].name
out = self._session.run(None, {inp_name: arr})[0]
kpts = _yolo_to_keypoints(out)
if _DEBUG_POSE:
print(f"[pose] raw_out={out.shape}, n_kpts_above_thresh={len(kpts)}")
pose_tags = _keypoints_to_pose_tags(kpts)
confs = [kp[:, 2].mean() for kp in kpts if kp.size]
pose_score = float(np.mean(confs)) if confs else 0.0
return {
"pose_tags": pose_tags,
"people_count": len(kpts),
"pose_score": round(pose_score, 4),
}
_pose_instance: PoseEstimator | None = None
def get_pose_tagger() -> PoseEstimator:
global _pose_instance
if _pose_instance is None:
_pose_instance = PoseEstimator()
return _pose_instance