ABot-Recon: Streaming 3D Reconstruction

Revisiting Local Context for Long-Horizon Streaming 3D Reconstruction

ABot-Recon reconstructs long video streams with a fixed 12-frame local context, composing current-frame geometry and adjacent relative poses into a global reconstruction without persistent learned long-range memory.

Usage

Upload an RGB video (MP4 recommended). The model extracts frames at the chosen interval, runs streaming inference, and produces:

• A colored 3D point cloud with rainbow camera trajectory overlay, shown in a custom WebGL viewer (drag to orbit, shift/right-drag to pan, scroll to zoom) with a ▶ Play trajectory button that flies the camera along the reconstructed path, plus a scrub bar and reset-view button
• A bird's-eye-view (BEV) trajectory plot
• A downloadable PLY point cloud

Model

• Checkpoint: acvlab/ABot-Recon
• Code: github.com/amap-cvlab/ABot-Recon
• Paper: arXiv:2608.27529

Notes

• Loop closure is disabled in this demo for faster inference.
• The model uses SDPA attention (FlashInfer-free) for ZeroGPU compatibility.
• Maximum 200 inference frames per run to stay within ZeroGPU time budget.
• Example video demo0.mp4 courtesy of the ABot-Recon authors (Apache-2.0 license).
• The 3D view is a custom HTML/JS (three.js) component rather than the default gr.Model3D viewer, so the reconstructed camera trajectory can be played back as an animated fly-through.