Fast3R Experimenter

A Python CLI tool to run Fast3R over single or multiple scenes with timings, PnP poses, and coloured fused point clouds, with optional view‐dropping and pointmap‐head choice.

Requirements

• Python 3.7+
• PyTorch
• numpy
• Pillow
• Open3D
• fast3r (and its dust3r dependencies)

Installation

pip install torch numpy Pillow open3d

Please visit the official Fast3r implementation repo for installation instructions.

Usage

Single Scene

python fast3r_experimenter.py \
  --images              /path/to/scene/images \
  --workspace           /results/scene \
  --size                512 \
  --dtype               fp32 \
  --pointmap_head       local \
  --conf_threshold      0 \
  --per_view_percentile 0

Batch

python fast3r_experimenter.py \
  --dataset_root        /data/multi_scenes \
  --workspace           /results/all \
  --size                512 \
  --dtype               bf16 \
  --pointmap_head       global \
  --conf_threshold      0 \
  --per_view_percentile 0

Arguments

Flag	Required?	Default	Description
--images	either this or --dataset_root	—	Folder containing your scene images (single-scene mode)
--dataset_root	either this or --images	—	Parent folder of scene subdirectories (batch mode)
--workspace	yes	—	Where to write per-scene outputs
--size	no	512	Resize longer image edge to this value
--dtype	no	fp32	Inference data type: fp32 or bf16
--pointmap_head	no	local	Which point-map head to fuse: local (uses local_pts3d_global_est) or global (falls back to pts3d_in_other_view)
--conf_threshold	no	0	Per-point confidence cutoff (points with conf <= threshold are dropped)
--per_view_percentile	no	0	Drop bottom P% of views by mean confidence (views with mean conf below this percentile are skipped)

Outputs (per scene)

• fused_fast3r.ply
  Coloured fused point cloud (if any points remain after masking).

• poses_c2w.npy
  NumPy array of camera-to-world poses.

• timings.csv
  CSV listing elapsed seconds for each pipeline stage (load_model, load_imgs, inference, dropped_views, align_pts, pose_pnp, fuse_pts, total).

• Console summary
  Prints number of views processed, total 3D points, and mean distance of points.