When dance feedback must survive the connection

Browser inference wins when dance feedback must continue through a network interruption without showing poses older than 120 ms. A server model can improve recognition, but its accuracy becomes irrelevant once its output misses that freshness limit. The architecture decision turns on whether silence during disconnection is acceptable.

Where the decision flips

Compare a browser that runs pose estimation and movement classification locally with a browser that sends captured frames to a remote inference service. Both must meet the same recognition acceptance criteria on the target device and dance vocabulary. If the application may suspend correction when the connection disappears, remote inference remains a valid choice when its measured recognition advantage warrants the dependency. If correction must continue beyond the freshness limit, local inference wins, provided the device sustains the complete feedback path.

A frozen skeleton is not continuity

When the connection fails, keeping the last skeleton visible turns an old observation into an apparently current correction, so stale output must lose its authority to guide the dancer.

Keep the dependency outside the loop

For the local architecture, capture, pose estimation, temporal classification and feedback rendering must operate without a remote response. Apply the W3C media capture specifications to camera access, while treating recovery and output freshness as application responsibilities. Load the model and required assets before the session, and keep remote analytics outside the feedback dependency chain. Carry the capture timestamp through processing, discard superseded frames and invalidate feedback when its source observation exceeds the freshness limit. This bounds the age of displayed evidence, although it does not guarantee that the classifier recognises the movement correctly.

A fallback must already be ready

A remote architecture with a local fallback only preserves recognition if the local model is ready and its temporal state is usable when the connection disappears. Starting it after the interruption creates a gap, while maintaining it during the session consumes device resources and reduces the benefit of offloading. Before design freeze, interrupt the connection during representative choreography and verify both feedback freshness and classification continuity on supported devices. Choose remote inference only if that gap is acceptable or the prepared fallback passes the same acceptance criteria, otherwise reserve enough device capacity for local inference to own the live session.

Back to DanceRecognition · Domains

← Back to DanceRecognition
Engineering & Architecture Brief

When dance feedback must survive the connection

Browser inference wins when dance feedback must continue through a network interruption without showing poses older than 120 ms. A server model can improve recognition, but its accuracy becomes irrelevant once its output misses that freshness limit. The architecture decision turns on whether silence during disconnection is acceptable.

Where the decision flips

Compare a browser that runs pose estimation and movement classification locally with a browser that sends captured frames to a remote inference service. Both must meet the same recognition acceptance criteria on the target device and dance vocabulary. If the application may suspend correction when the connection disappears, remote inference remains a valid choice when its measured recognition advantage warrants the dependency. If correction must continue beyond the freshness limit, local inference wins, provided the device sustains the complete feedback path.

A frozen skeleton is not continuity

When the connection fails, keeping the last skeleton visible turns an old observation into an apparently current correction, so stale output must lose its authority to guide the dancer.

Keep the dependency outside the loop

For the local architecture, capture, pose estimation, temporal classification and feedback rendering must operate without a remote response. Apply the W3C media capture specifications to camera access, while treating recovery and output freshness as application responsibilities. Load the model and required assets before the session, and keep remote analytics outside the feedback dependency chain. Carry the capture timestamp through processing, discard superseded frames and invalidate feedback when its source observation exceeds the freshness limit. This bounds the age of displayed evidence, although it does not guarantee that the classifier recognises the movement correctly.

A fallback must already be ready

A remote architecture with a local fallback only preserves recognition if the local model is ready and its temporal state is usable when the connection disappears. Starting it after the interruption creates a gap, while maintaining it during the session consumes device resources and reduces the benefit of offloading. Before design freeze, interrupt the connection during representative choreography and verify both feedback freshness and classification continuity on supported devices. Choose remote inference only if that gap is acceptable or the prepared fallback passes the same acceptance criteria, otherwise reserve enough device capacity for local inference to own the live session.