Skip to content

Speed up Base depth and preserve Windows reconstruction preview - #196

Open
REDrighthand05 wants to merge 1 commit into
mainfrom
perf/windows-reconstruction-main
Open

REDrighthand05 wants to merge 1 commit into
mainfrom
perf/windows-reconstruction-main

Conversation

@REDrighthand05

Copy link
Copy Markdown
Owner

Relates to #74. Uses the v0.6.81 payload from #195 and carries forward #194’s WebGL reuse and atomic streamed scene output for the current desktop code.

  • Base uses one CPU ORT session by default. GPU depth keeps ordered, bounded preprocessing, respects fixed batch shapes and retries rejected dynamic batches without reprocessing their frames. The job records phase timing and evidence of CUDA model-node execution.
  • Native caption controls keep their existing path. The page coalesces resize requests, reuses WebGL state and receives preview images only when visible and newer.
  • TUM Freiburg1 desk RGB, the same 32-frame MP4 and release Base model on Linux CPU: 108.13 → 72.20 s end to end. 32/32 frames registered, inferred, and fused on both runs. All depth maps matched bit-for-bit. The independently estimated tracks varied slightly.

Input hashes, stage timings and Windows/NVIDIA reproduction commands. The measured gain is for this Linux CPU sample. Windows installer frame latency and CUDA end-to-end speed need device validation.

@REDrighthand05
REDrighthand05 marked this pull request as ready for review November 19, 2026 13:53
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant