Where SLS came from
Structured Light Sensor cameras (originally the Microsoft Kinect) project a pattern of infrared dots into a room and read how that pattern deforms to build a depth map. They were trained to find human shapes so a game could detect a player. SpiritMesh uses pose-detection neural networks that work on regular video, but the principle is similar: the model is trying to fit a skeleton to anything that looks vaguely person-shaped.
What it actually detects
Joints — head, shoulders, elbows, hips, knees, ankles. The model emits a confidence score for each joint. A fully-formed humanoid figure usually requires 8+ joints with average confidence >30%. Fewer joints, lower confidence, or rapidly flickering joints generally mean something humanoid-shaped, not a person.
Common false-figures
Coat racks. Chairs with throws on them. Tall plants. Curtains catching air. Your own reflection in a dark window. The model will happily 'see' a person in any of these — that's why depth stability and motion coherence matter as much as joint count.
Reading the three numbers
Pose confidence = how sure the model is of joint positions. Depth stability = how consistent the figure is in space across frames. Motion coherence = whether the figure moves like a coordinated body or like noise. A clip with high pose but low depth/motion is almost always a static object being misread.