When a computer vision model looks at a human, it sees a collection of coordinates. However, these coordinates are highly sensitive to "nuisance variables": how far the person is from the lens, the angle of the camera, and the person's unique physical proportions. To a raw neural network, a tall person walking toward a camera and a short person walking across the frame look like entirely different mathematical entities, even if they are performing the exact same action.
Anatomical scale and coordinate normalization is the process of stripping away these environmental and biological variables to reveal the underlying "geometry" of movement. By transforming raw skeletal data into a standardized, human-centric reference frame, researchers enable models to generalize across different subjects, cameras, and datasets.
Eliminating Global Translation and Rotation
The first step in understanding a pose is to ignore where the person is in the world and focus on where their limbs are relative to their body. Most frameworks begin by "centering" the skeleton. Fan et al. and Babey et al. demonstrate that by setting a central anatomical point—typically the pelvis or hip center—as the origin (0,0,0), the model becomes invariant to the subject's absolute position in the room.
However, centering is not enough if the subject is rotated. Shrivastava et al. and Demisse et al. use algorithms like Kabsch alignment to rotate 3D joints into a fixed coordinate frame. This "canonical pose normalization" ensures that if two people are both raising their right arms, their skeletal vectors point in the same direction in the model's latent space, regardless of whether one person is facing North and the other South. Pala et al. even suggest using Principal Component Analysis (PCA) to find the skeleton's internal axes, ensuring the representation is defined by the body's own symmetry rather than the camera's perspective.
Decoupling Scale from Motion
A significant challenge in biometric analysis, such as gait recognition, is that a person's stride is inextricably linked to their height. To compare the way someone walks rather than how big they are, researchers must normalize for scale.
An et al. and Malin-Mayor et al. achieve this by scaling the entire skeleton based on the distance between stable anatomical landmarks, such as the neck and the hip. This ensures that a child and an adult performing the same movement produce identical normalized features. Taking this further, Hakim proposes enforcing uniform bone lengths across all subjects using a graph traversal method. By making every "digital human" have the exact same limb lengths, the model is forced to learn motion patterns—the change in angles over time—rather than being distracted by physical stature.
Local Anchoring for Fine-Grained Interaction
While a pelvis-centered frame works for full-body actions, it can be too coarse for tasks involving hands or objects. In these cases, researchers shift the "anchor" to the specific body part performing the action.
For sign language recognition, Hosain and Medronha et al. transform coordinates to be relative to the wrist. This spatial normalization allows the model to focus on the intricate geometry of the fingers rather than the arm's position in space. Similarly, Gupta et al. show that when reconstructing objects held in a hand, treating the wrist as the local origin makes the system robust to camera perspective, as the relationship between the hand and the object remains constant even if the person moves their arm.
Unifying Skeletal Topology
A final hurdle in this field is "dataset bias." Different motion capture systems and datasets often use different skeletal hierarchies (e.g., one might track 17 joints, another 25).
To bridge this gap, Bensabath et al. map diverse data sources to a standardized body model format, ensuring consistent joint configurations. Yu et al. take a more radical approach by using linear interpolation to sample a fixed number of points along body segments. This effectively "resamples" the human body into a unified input format, allowing a single Transformer model to process motion data regardless of the original sensor's configuration.
By combining these techniques—centering, rotating, scaling, and re-topologizing—researchers transform raw, noisy video data into a "mathematical skeleton" that is universal. As Fu et al. demonstrate, this normalization can improve cross-domain recognition accuracy by over 30%, proving that the key to understanding human action is first understanding how to ignore the environment in which it happens.
Key Foundational Work
Action Recognition with Exemplar Based 2.5D Graph Matching (2012) — contributed 2 concept elements to this direction.