[Bug] Gaussian point clouds are not scaled during metric depth alignment, causing pose inconsistency

[Phrandime]

Problem Description

In NestedDepthAnything3Net._apply_depth_alignment(), when converting predicted depth to metric scale, the code scales depth and extrinsics but misses scaling the gaussians. This causes the exported Gaussian point clouds (via gs_ply) to have a different scale than the camera poses (via cameras.json), resulting in incorrect rendering when using standard 3DGS viewers.

Root Cause

In src/depth_anything_3/model/da3.py, lines 410-414:

# Apply scaling to depth and extrinsics
output.depth *= scale_factor
output.extrinsics[:, :, :3, 3] *= scale_factor
output.is_metric = 1
output.scale_factor = scale_factor.item()

The output.gaussians (which contains means and scales) is not scaled, even though it was computed using the pre-scaling depth values.

Impact

When exporting to 3DGS format (--export-format gs_ply):

• Before fix: Gaussian points have ~40x larger extent than camera positions
• Result: Rendering from predicted camera poses shows completely wrong views (PSNR ~7 dB)
• Verification: 3DGS viewers cannot correctly display the scene from the exported cameras

Proposed Fix

Add scaling for gaussians in _apply_depth_alignment():

# Apply scaling to depth and extrinsics
output.depth *= scale_factor
output.extrinsics[:, :, :3, 3] *= scale_factor
output.is_metric = 1
output.scale_factor = scale_factor.item()

# FIX: Also scale gaussians if they exist to maintain consistency
if hasattr(output, 'gaussians') and output.gaussians is not None:
    output.gaussians.means *= scale_factor
    output.gaussians.scales *= scale_factor

Files Affected

• src/depth_anything_3/model/da3.py: _apply_depth_alignment() method (around line 416)

Additional Context

This issue only affects users who:

1. Use the nested/metric model (DA3NESTED variants)
2. Export to 3DGS format (gs_ply)
3. Attempt to render from the exported camera poses

The scene.glb export is not affected because it uses COLMAP point clouds which are computed from the scaled depth maps.