Pose Estimation and OBB
Two specialized tasks and the kinds of problems where they shine.
The core tasks — classification, detection, instance segmentation, and semantic segmentation — handle most CV work. The rest is where pose estimation and OBB live. Both are easy to dismiss until you hit the exact problem they were made for — like workouts monitoring or aerial inspection — and then nothing else works.
Hands-on
Pose estimation#
A pose model returns a list of (x, y) keypoints per detected subject — usually a person, but it works for animals or rigid objects too. The COCO-Pose Person model has 17 keypoints (eyes, ears, shoulders, elbows, wrists, hips, knees, ankles), and YOLO26 pose is the recommended Ultralytics default.
You don't use pose to answer "is there a person here?" — that's a detection question. You use it when the answer depends on what the person is doing, a form of action recognition:
- Fall detection: a body lying horizontally with low keypoint variance.
- Workplace ergonomics: how often a worker bends at the back rather than the knees.
- Sport analysis: stride length, swing angle, jump height.
- Animal welfare: posture changes that suggest distress in livestock.
- Sign detection / hand tracking: finger positions over time.
results = model("athlete.jpg")
keypoints = results[0].keypoints.xy # tensor: [n_people, 17, 2]
for person in keypoints:
nose, left_shoulder, right_shoulder = person[0], person[5], person[6]
print("nose:", nose.tolist(), "shoulders:", left_shoulder.tolist(), right_shoulder.tolist())A pose model still returns boxes — it just adds keypoints inside each box. If you don't need the keypoints, a vanilla detector is faster and cheaper.
Oriented bounding boxes (OBB)#
A regular box has four numbers: corners. An OBB has five: center, width, height, and angle. That extra angle is the whole point — see the DOTAv2 dataset for canonical aerial OBB examples.
OBB shines when objects don't sit parallel to the image frame:
- Aerial / satellite imagery: ships, aircraft on tarmacs, vehicles in parking lots — most are at random angles when seen from above.
- Top-down conveyor systems: packages oriented however they fell on the belt.
- Document analysis: photographed text and forms that aren't perfectly aligned.
- Inventory in warehouses: crates and pallets at angles to the camera.
The reason axis-aligned detectors struggle here isn't accuracy — it's that the box hull badly overstates the object. A 30-meter ship at 45° fills a box twice the area of the ship itself, with most of the box being water. NMS can't separate two ships if their hulls overlap by 60% — even if the actual ships don't touch.
| Use case | Axis-aligned | OBB |
|---|---|---|
| Counting people on a sidewalk | ✅ | overkill |
| Counting docked ships from above | ⚠️ NMS confuses neighbors | ✅ |
| Detecting cars in a city street | ✅ | overkill |
| Detecting cars in a parking lot from above | ⚠️ | ✅ |
| Reading rotated invoices | ⚠️ | ✅ |
Choosing a specialized task#
The decision is usually clear once you ask the right question:
- Is the angle of the object meaningful? → OBB.
- Is the posture of the subject the prediction? → Pose.
- If neither — stick with detection or segmentation, optionally with object tracking for cross-frame identity.