Block Calibration for Open-Vocabulary Detection

Code and data for Deciding Which Detections to Release: Block Calibration for Open-Vocabulary Detection.

Block calibration keeps each image and its queries together, then applies e-BH to decide which detections to release. Under exchangeable image blocks and a fixed query rule, it controls expected query-average FDP while allowing score-dependent candidate counts and dependence within an image.

Quick start

Python 3.10+; the cached experiments run on CPU without detector weights.

python -m venv .venv
source .venv/bin/activate
pip install -r requirements-paper.txt
python scripts/reproduce_block_results.py --verify-data

This replays all ten COCO/GroundingDINO-tiny splits and checks them against the archived measurements. At alpha=0.025, learned evidence gives 16.51% object recall / 98.12% pooled precision, versus 8.07% / 99.03% for binary evidence. Add --all-settings to replay all seven detector–dataset settings.

Contents

• paper/main.pdf: current manuscript; LaTeX sources, figures and tables are in paper/.
• paper/experiments/: current block-calibration implementation, baselines and replay protocol.
• paper/source_data/: figure data, per-split results and release records.
• data/: compressed caches for 70 image splits, prompt-bank inputs and checksums.
• REPRODUCING.md: full replay and figure commands, data provenance and scope.

The pre-existing outputs/, src/ and other scripts retain historical candidate-pooled studies; use the entry points above for the current paper.

Raw benchmark images and detector weights are obtained separately from COCO, LVIS, Open Images V5 and the model providers. Dataset and model licenses continue to apply. Project code is MIT licensed; third-party figure notices are retained.

Build the paper with cd paper && latexmk -pdf main.tex.