Dream2Detect

Dream2Detect studies synthetic-to-real generalization for visible package/cardboard defect severity.

Presentation: https://canva.link/4wm0l3yeq6bvidc

Report: https://docs.google.com/document/d/1DHi4dMjjbmDWHcfkbAPDVu5CtHloaJb2uUcFfvDqQ_k/edit?usp=sharing

Teacher Quick View

• Start here: DELIVERY.md
• Pipeline summaries: pipelines/README.md
• Main pretrained model notes: pipelines/v0-main-pretrained/FinalModels/README.md

Layout

dataset/               compact delivery datasets, ready to run
dataset_generation/    preserved prompt/image/QC generation workflow
pipelines/             one folder per pipeline, each documented
checkpoint_tools/      teacher-facing scripts for testing saved checkpoints
scripts/               shared runnable entrypoints for V1-V5B
src/dream2detect/      minimal shared training package for V1-V5B

Delivery Datasets

The delivery branch includes two compact dataset folders:

• dataset/synthetic/manifest.csv
• dataset/real/manifest.csv

Each manifest uses image_path values relative to its own dataset folder, for example images/example.png.

Validate them with:

python3 scripts/validate_delivery_datasets.py

Common Commands

python3 scripts/train_synthetic_classifier.py \
  --manifest dataset/synthetic/manifest.csv \
  --image-size 384 \
  --output-dir outputs/v4_score_band

python3 scripts/train_synthetic_regressor.py \
  --manifest dataset/synthetic/manifest.csv \
  --output-dir outputs/v5_a_scalar

python3 scripts/train_multitask_classifier.py \
  --manifest dataset/synthetic/manifest.csv \
  --output-dir outputs/v5_b_multitask

Test Saved Checkpoints

Place trained .pt files under checkpoints/, then run:

PYTHONPATH=.:src python3 checkpoint_tools/test_checkpoints.py \
  --checkpoints-dir checkpoints \
  --manifest dataset/real/manifest.csv \
  --output-dir checkpoint_results