Day 56: Pose estimation landscape: top-down vs bottom-up
Finding the skeleton
With the person located, the next component is pose estimation: locate anatomical keypoints — shoulders, elbows, wrists, hips, knees. Connecting them gives a skeleton, and the distances between keypoints (shoulder width, arm length, torso height) are the pixel measurements that, once calibrated, become the centimeters the Measurement Engine reports.
Top-down vs bottom-up
- Top-down — first detect each person (you already have this from YOLO), then estimate keypoints within each person's box. Accurate; cost scales with number of people.
- Bottom-up — detect all keypoints in the image at once, then group them into people. Faster for crowds; trickier grouping.
- For FitXpert (usually one person, front-on), top-down fits naturally — your detector already provides the person box.
FitXpert's constraint simplifies the problem
Because FitXpert photos are typically a single person standing front-on, you avoid the hardest pose problems (crowds, heavy occlusion, extreme poses). Recognizing which hard problems your product *doesn't* have is a senior instinct — it lets you pick the pragmatic tool (MediaPipe, tomorrow) instead of over-engineering.
Key terms
- Pose estimation
- Locating anatomical keypoints (joints) of a person in an image.
- Keypoint
- A specific body landmark (e.g. left shoulder, right hip) predicted by a pose model.
- Top-down pose
- Detecting each person first, then estimating keypoints within each detected box.
Why does top-down pose estimation fit FitXpert naturally?
