README.md

Experiments with the detector-retrieval baseline

The detector-retrieval baseline consists of the two models that need to be trained independently.

Installation

See INSTALL.md

Preparations

conda activate os2d
export PYTHONPATH=$OS2D_ROOT:$PYTHONPATH

Train the detector

The main training script is detector/train_detector.py. The training runs can be launched with detector/launcher_train_detector.py.

cd $OS2D_ROOT/baselines/detector_retrieval/detector

#6 jobs training class agnostic detectors, 2 jobs for training class-aware detectors
python experiments/launcher_train_detector.py

Trained model will be stored in $OS2D_ROOT/baselines/detector_retrieval/detector/output.

Train the retrieval system

Prepare dataset

One needs to convert datasets to the format of cnnimageretrieval-pytorch. This csript should do this:

cd $OS2D_ROOT/baselines/detector_retrieval/retrieval

bash prepare_all_datasets.sh

Run training

cd $OS2D_ROOT/baselines/detector_retrieval/retrieval

# Grozi
python experiments/launcher_grozi.py
# INSTRE
python experiments/launcher_instre.py

Run evaluation

cd $OS2D_ROOT/baselines/detector_retrieval

# Grozi
python experiments/launcher_grozi_eval.py
# INSTRE
python experiments/launcher_instre_eval.py

View results

cd $OS2D_ROOT/baselines/detector_retrieval

# Create tables
# Table 3
python experiments/launcher_grozi_eval_collect.py
# Table 4
python experiments/launcher_instre_eval_collect.py