Pose Estimation and OBB

Pose: when posture is the signal#

A pose model returns a box for each person plus a fixed list of keypoints. Models trained on COCO-Pose predict 17 per person — nose, eyes, ears, shoulders, elbows, wrists, hips, knees, and ankles — always in the same order, so index 0 is the nose, 11 the left hip, and 15 the left ankle. Each keypoint also has a confidence.

The simplest useful rule: a standing body is taller than it is wide; a body lying on the floor is wider than it is tall. Here it is on the bus image, using only keypoints the model is confident about:

from ultralytics import YOLO

model = YOLO("yolo26n-pose.pt")
result = model("https://ultralytics.com/images/bus.jpg", verbose=False)[0]

for xy, conf in zip(result.keypoints.xy, result.keypoints.conf):
    visible = xy[conf > 0.5]  # drop keypoints the model couldn't see
    width = float(visible[:, 0].max() - visible[:, 0].min())
    height = float(visible[:, 1].max() - visible[:, 1].min())
    posture = "unknown" if len(visible) < 5 else "lying" if width > height else "upright"
    print(f"{len(visible):2d}/17 keypoints visible, body {width:3.0f} px wide x {height:3.0f} px tall -> {posture}")

16/17 keypoints visible, body 117 px wide x 427 px tall -> upright
 6/17 keypoints visible, body  80 px wide x 359 px tall -> upright
16/17 keypoints visible, body  91 px wide x 384 px tall -> upright
 1/17 keypoints visible, body   0 px wide x   0 px tall -> unknown
 2/17 keypoints visible, body   3 px wide x   6 px tall -> unknown

Three people are clearly upright. The last two rows are the person cut off at the left edge and a low-confidence duplicate: with one or two visible keypoints there's nothing to judge, so the rule says "unknown" instead of guessing. A production fall detector adds time — lying for more than a few seconds, after a fast drop — through tracking.

Pose earns its place when the answer depends on what the body is doing: falls or lifting with a bent back for safety, joint angles and rep counts in sport as in workouts monitoring, or posture in livestock and fixed landmarks on parts with a custom keypoint layout. To answer "is there a person here?", use plain detection.

OBB: when angle matters#

A regular box has edges parallel to the image. An oriented bounding box adds a rotation, stored as center, width, height, and angle (xywhr). In top-down imagery — satellite, drone, overhead conveyor cameras — objects lie at every angle, and an upright box around a diagonal object is mostly background, as the yachts above would be. Measure how much on a drone shot of a marina:

from ultralytics import YOLO

model = YOLO("yolo26n-obb.pt")
result = model("https://ultralytics.com/images/boats.jpg", verbose=False)[0]

obb = result.obb
rotated_area = obb.xywhr[:, 2] * obb.xywhr[:, 3]  # width x height of each rotated box
x1, y1, x2, y2 = obb.xyxy.T  # the axis-aligned box that would enclose the same object
upright_area = (x2 - x1) * (y2 - y1)

print(f"{len(obb)} objects, classes: {sorted({result.names[int(c)] for c in obb.cls})}")
cx, cy, w, h, angle = obb.xywhr[0].tolist()
print(f"first box: center=({cx:.0f}, {cy:.0f}) size={w:.0f}x{h:.0f} angle={angle:.2f} rad")
ratio = float((upright_area / rotated_area).median())
print(f"an axis-aligned box would be {ratio:.1f}x larger (median)")

173 objects, classes: ['harbor', 'ship']
first box: center=(619, 850) size=33x108 angle=-0.25 rad
an axis-aligned box would be 2.1x larger (median)

An upright box covers about twice the area of the rotated box around the same boat, and the extra area is water. It hurts twice: NMS sees the upright boxes of neighboring boats overlapping and may suppress real ones, and any size estimate is inflated. OBB models trained on aerial datasets such as DOTA handle these scenes directly. Note the angle is in radians: -0.25 rad is about -14°.

Choosing between detection, pose, and OBB#