A surveillance camera captures a figure in a heavy winter coat. The silhouette is a shapeless blob. Background subtraction struggles. The coat hides the trunk, merges the legs, and buries the posture signature we spent the last lecture building. A silhouette-based system falters. But a model-based system does something different. It looks past the coat entirely. It asks: where are the joints? Model-based approaches overcome the limitations of silhouette quality by focusing on joint locations and movement trajectories, making them less susceptible to shadows, clothing, and viewpoint variations. Model-based gait recognition takes a different path. Instead of reading the clothing-wrapped outline, it represents a person as a sequence of body-joint locations. A mathematical skeleton. The coat becomes irrelevant because the system is tracking the geometry underneath it. Think of the pipeline as a five-step assembly line. A typical pipeline detects the person in the frame. Second, estimate pose: predict the locations of hips, knees, ankles, shoulders, elbows, and wrists. Third, track that skeleton across time. Fourth, normalize it. Fifth, extract motion features and classify identity. Each step feeds the next. Each step can also break the next. That chain dependency helps the system work, and it also creates a serious vulnerability. The digital skeleton is a geometric abstraction that relies on accurate pose estimation, which can be affected by camera view, image quality, and occlusion. What it preserves is temporal movement: how each joint changes position across successive frames. That sequence of positions becomes the biometric signal. Common features include joint coordinates, inter-joint distances, limb lengths, joint angles, and joint velocities. The key idea is that identity lives in the trajectory, not the snapshot. Raw joint coordinates carry noise from camera distance and frame position. Translation normalization recenters the skeleton around a reference joint, typically the pelvis. Scale normalization reduces the influence of camera distance, though it may also strip out useful anthropometric information. Rotation normalization attempts to express motion in a common coordinate system across viewpoints. Then graph-based models take over: joints become nodes, anatomical connections become edges, and the model learns relationships between body parts across time. Here is the counterintuitive part, Jordan. A simplified stick-figure skeleton can outperform a rich image-based representation, because it discards appearance noise. But estimated limb length may still leak anthropometric identity even after normalization. And pose-estimation errors propagate hard. One study found a Pearson correlation of 0.919 between pose-estimator performance and downstream recognition accuracy. That means upstream errors are not absorbed by the system. They amplify. Incorrect joint locations create artificial stride patterns that the classifier reads as identity. Despite the promise of clothing invariance, skeleton-based systems can still experience accuracy drops, as seen in the CASIA-B benchmark where accuracy fell significantly when subjects wore coats. That twenty-point gap is striking, Jordan, because the whole promise of skeleton methods is clothing invariance. The coat changes the visible silhouette, which changes the pose estimate, which corrupts the skeleton. The chain breaks at step two. That means a practical system must evaluate cross-view, cross-clothing, and cross-session conditions separately, not as one blended accuracy number. [short pause] Model-based gait recognition maps the movement of specific joints, creating a mathematical skeleton of the subject. That skeleton reduces sensitivity to clothing and background. But its performance depends heavily on the pose estimator feeding it. The takeaway is this: the digital skeleton is a powerful abstraction, but abstraction has a cost. Estimation errors upstream can become recognition errors downstream. The walk can reveal the architecture beneath it when the system can find that architecture in the video.