https://jaypolra.github.io/tags/cvpr-2026/CVPR 2026Hugo Blox Builder (https://hugoblox.com)en-usMon, 01 Jun 2026 00:00:00 +0000https://jaypolra.github.io/media/icon_hu_3943bc2ea2cd9fa1.pnghttps://jaypolra.github.io/tags/cvpr-2026/https://jaypolra.github.io/publication/nvas-cvpr-2026/Mon, 01 Jun 2026 00:00:00 +0000https://jaypolra.github.io/publication/nvas-cvpr-2026/https://jaypolra.github.io/project/geometry-grounded-nvas/Thu, 01 Jan 2026 00:00:00 +0000https://jaypolra.github.io/project/geometry-grounded-nvas/<p>Prior audio-visual methods depend on Structure-from-Motion to build geometry: expensive, brittle with sparse frames, and unavailable at inference in many real scenes.</p> <p>Built a feed-forward pipeline using VGGT geometry encoding and a cross-attention decoder that routes on geometry but retrieves learned acoustic priors. No COLMAP, no target-view image required.</p> <p>Key architectural decision: separate keys and values in cross-attention so visual geometry decides which reference views are relevant, but the actual retrieved information is the learned acoustic fingerprint of those views, not the visual features themselves.</p> <p>Outperformed prior SOTA (AV-Cloud) on all four metrics (MAG, ENV, LRE, DPAM) with fewer parameters (3.24M vs 3.91M) and 10x faster preprocessing than COLMAP.</p>