Dataset Creation

Utilities for generating synthetic workplace datasets with Gemini image generation and optional Roboflow-assisted bounding boxes.

The repository has two independent workflows:

• room_readiness_generation/: creates paired room-state images for whiteboards, chairs, blinds, and tables.
• participant_classification_generation/: edits base workplace images to add meeting participants and non-participants, then optionally generates and previews person/head bounding boxes.

Generated images, labels, previews, API keys, virtual environments, and local cache files are intentionally ignored by git. A fresh clone should contain code and documentation only.

Setup

Use Python 3.11 or newer.

python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
cp .env.example .env

Edit .env with the API keys you need:

• GEMINI_API_KEY for Gemini image generation.
• ROBOFLOW_API_KEY for Roboflow inference.

Most scripts also read exported environment variables, so this works too:

export GEMINI_API_KEY="..."
export ROBOFLOW_API_KEY="..."

Room Readiness Workflow

Run from the repository root:

cd room_readiness_generation
python generate_base_images.py
python main.py
python change_format.py

Outputs:

• room_readiness_generation/base_images/: generated base rooms.
• room_readiness_generation/output/: paired variant images plus labels and cost logs.
• room_readiness_generation/new_outputs/: flattened category folders with annotations.json.

Useful dry-run commands:

python generate_base_images.py --dry-run
python main.py --dry-run
python main.py --subtask chairs --dry-run

Interactive review UI:

python app.py

Then open the local Flask URL printed by Flask, usually http://127.0.0.1:5000.

Participant Classification Workflow

Put source images in participant_classification_generation/base_images/, then run:

cd participant_classification_generation
python datasetGenScript.py --prompt most_non --max-images 50

Prompt choices:

• no_non: all visible people should be meeting participants.
• some_non: a few visible non-participants are present.
• most_non: many visible non-participants are present.

To generate automatic boxes with Roboflow, start the local inference server first:

inference server start

Then run:

python roboflow_autoBounding.py \
  --image-folder generated_images \
  --output-json auto_boxes.json

python visualizeBoxes.py \
  --image-folder generated_images \
  --json-file auto_boxes.json \
  --output-folder preview_boxes

Reproducibility Notes

• Keep prompt edits in source files so future runs are explainable.
• Keep generated datasets out of git; publish large outputs through a dataset store or release artifact instead.
• Use --dry-run where available before spending API credits.
• room_readiness_generation/main.py writes token and cost metadata to output/cost_log.json.