VPhotoBench / evaluation /EVALUATION_PROTOCOL.md
sjijs's picture
Upload 2 files
40d9125 verified

Using UniPercept for VPhotoBench Outputs

VPhotoBench provides Blender scenes and text prompts. After running an agent or camera-search method, users can evaluate the final rendered images with UniPercept:

https://github.com/thunderbolt215/UniPercept

UniPercept returns three perceptual scores for each image:

  • IAA: image aesthetics assessment.
  • IQA: image quality assessment.
  • ISTA: image structure and texture assessment.

1. Prepare Results

Create a CSV file with one row per rendered result:

task_id,image_path
001_blender_4_0_splash_subject_placement,images/001_blender_4_0_splash_subject_placement.png

task_id should match the IDs in benchmark/tasks.json. image_path can be absolute or relative to the submission/result directory.

2. Install UniPercept

Follow the official UniPercept instructions. The lightweight package interface can be installed with:

pip install unipercept-reward

If you use a local checkpoint or a specific UniPercept release, follow the setup steps in the UniPercept repository.

3. Score Images

Run:

python evaluation/run_unipercept_scoring.py \
  --benchmark-root . \
  --submission path/to/submission.csv \
  --submission-root path/to/result_directory \
  --output path/to/unipercept_scores.csv \
  --device cuda

The output CSV contains:

task_id,scene_id,mission_type,image_path,status,iaa,iqa,ista,m_qs,succ_at_0_55,error

The script normalizes UniPercept scores from 0-100 to 0-1. It also reports an optional composite score:

M_qs = 0.40 * IAA + 0.20 * IQA + 0.40 * ISTA

and:

Succ@0.55 = 1 if M_qs >= 0.55 else 0

These two derived fields are provided only as a convenient default summary. Users may report the three UniPercept scores directly or define their own aggregation rule for a specific study.