Question about reproducing the 7-Scenes benchmark settings

Hi, thank you for releasing this cool work and the codebase.

I understand that an official evaluation benchmark may be released later,
but in the meantime I am trying to reproduce the 7-Scenes numbers reported in the paper.
In particular, I am looking at the 7-Scenes results in Table 4 and Table 5:

• Pose: AUC@3 = 12.63, AUC@30 = 78.59, ATE = 0.08
• Reconstruction: Acc = 0.02, Comp = 0.07, F1 = 80.39

With my current setup, I get lower numbers than the paper, so I suspect that my evaluation/inference protocol may not match yours.

The command I used is roughly:

python demo.py \
  --model_path ./models/lingbot-map.pt \
  --image_folder /path/to/7scenes/<scene>/<seq> \
  --image_ext ".color.png" \
  --stride 5 \
  --mode streaming \
  --image_size 518 \
  --patch_size 14 \
  --num_scale_frames 8 \
  --no_visualization \
  --save_predictions /path/to/output.pt

For reconstruction evaluation, I used voxel size 4.0 / 512, ICP threshold 0.1, and F1 threshold 0.05, following the paper description as closely as possible.

Could you share the approximate settings used to obtain the reported 7-Scenes results? In particular, I would like to confirm:

• Which checkpoint was used for Table 4/5: lingbot-map.pt, lingbot-map-long.pt, or another internal checkpoint?
• Was 7-Scenes evaluated in pure streaming mode, windowed mode, or with any state reset policy?
• What keyframe interval / cache policy / number of scale frames was used?
• Were the reconstruction metrics computed from the predicted world_points, or by unprojecting predicted depth with predicted poses?
• Are there any confidence filtering, sky/depth masking, pose convention, or alignment details that are important for matching the reported numbers?

Thanks again for the great work.

Start with demo.py and the reported Table 4 and Table 5 values, then compare the supplied command and reconstruction settings with the repository's documented evaluation path. Check the checkpoint, inference mode, state-reset and keyframe settings, metric inputs, and filtering or alignment details mentioned in the issue. Done means the settings needed to reproduce the reported 7-Scenes results are identified and documented.