PoseTracker API

Processes continuous video input (webcam, file, or streaming source) to detect and track a single human skeleton in real-time, outputting joint coordinates and confidence scores for 17-25 keypoints (depending on model variant). Uses deep neural network inference (likely convolutional backbone with heatmap regression or keypoint detection heads) optimized for low-latency inference on consumer hardware. Operates on standard RGB frames without requiring depth sensors, IR markers, or specialized capture equipment.

Returns per-joint confidence scores (typically 0.0–1.0) indicating model certainty for each detected keypoint, enabling developers to filter or weight unreliable detections. Confidence reflects the neural network's activation strength at that joint location and implicitly encodes uncertainty from occlusion, motion blur, or ambiguous body configuration. Developers can threshold confidence to discard low-quality keypoints before downstream processing (animation, physics, analytics).

Processes video frame-by-frame and outputs pose data for each frame with timestamps, enabling temporal analysis and motion reconstruction. Each frame produces a complete skeleton snapshot (all joint positions and confidences at that moment), allowing developers to compute velocity, acceleration, and motion patterns over time. Output is typically JSON arrays indexed by frame number or timestamp, preserving frame-to-frame correspondence for animation playback or motion analysis.

Exposes HTTP endpoints accepting video frames or file uploads, returning pose data in JSON format. Likely supports multiple model variants (e.g., lightweight for mobile, high-accuracy for desktop) selectable via query parameters or request headers. Inference runs server-side, abstracting model loading and GPU management from the client. Responses include pose keypoints, confidences, and metadata (model version, inference time, frame dimensions).

Provides free tier access to pose estimation with unspecified monthly or daily request limits, enabling developers to experiment and prototype before committing to paid plans. Quota enforcement likely implemented via API key rate limiting (requests per minute/hour) and monthly request caps. Freemium tier may have reduced model accuracy, longer inference latency, or lower priority in server queue compared to paid tiers.

Outputs pose keypoint data in formats compatible with animation tools (e.g., BVH, FBX, or proprietary game engine formats). Converts skeletal joint coordinates from PoseTracker's native representation into industry-standard motion capture formats, enabling direct import into Maya, Blender, Unreal Engine, or Unity. Likely includes bone hierarchy mapping, coordinate system transformation (e.g., Y-up to Z-up), and optional frame interpolation for smooth playback.

Analyzes sequences of pose frames to recognize high-level gestures or motion patterns (e.g., 'jumping', 'waving', 'squatting') by matching joint trajectories against learned pattern templates. Likely uses temporal convolution or hidden Markov models to classify motion sequences, outputting gesture labels with confidence scores. Enables applications to respond to user actions (e.g., 'user performed a squat') rather than raw joint coordinates.

Optimizes inference pipeline for minimal end-to-end latency (capture → inference → output), targeting interactive use cases like live gaming or VR. Likely employs model quantization (INT8), pruning, or distillation to reduce computational cost, and may support edge deployment (on-device inference) for sub-50ms latency. Streaming inference mode processes frames as they arrive without buffering, enabling responsive pose-driven interactions.