SpatialTunnel Dataset Generation

Blender Setup

Rendering is done through Blender's bundled Python. Install Blender on Linux by following the official manual:

https://docs.blender.org/manual/en/latest/getting_started/installing/linux.html

Use an installation method that makes the blender executable available on PATH. For a downloaded Linux tarball, this usually means extracting Blender and adding the extracted directory to PATH, or creating a symlink from the installed blender executable into a directory already on PATH.

Verify the command before running the tests:

command -v blender
blender --version

The current code is checked with:

Blender 5.1.1 (build date 2026-04-20)

Phase Variation

Run:

./run_phase_variation.sh

This executes:

blender -b spatialtunnel.blend \
  -P phase_variation.py -E CYCLES -- \
  --cycles-device CUDA \
  --outdir //output/phase_variation \
  --num 12

Behavior:

• Randomizes object color/shape once per variant.
• Varies Obj1 and Obj2 phases across 16 values each.
• Produces 256 renders per variant.
• Uses NUM=12, so the full run produces up to 3,072 images.
• Writes render outputs and log.csv under output/phase_variation.
• Writes stdout/stderr to output_phase_variation.log.

Size Variation

Run:

./run_size_variation.sh

This executes:

blender -b spatialtunnel.blend \
  -P size_variation.py -E CYCLES -- \
  --cycles-device CUDA \
  --outdir //output/size_variation \
  --num 100

Behavior:

• Randomizes object color/shape once per variant.
• Varies Obj1 size from 0.1 to 0.3.
• Sets Obj2 size to 0.4 - Obj1 size.
• Uses 10 size segments, producing 11 renders per variant.
• Uses NUM=100, so the full run produces up to 1,100 images.
• Writes render outputs and log.csv under output/size_variation.
• Writes stdout/stderr to output_size_variation.log.