DVAL: Detector and Vision-language active learning for traffic sign detection

This repository contains an active-learning pipeline for object detection with YOLO and MaPLe-based visual re-ranking. The current public setup is focused on DVAL experiments for two traffic-sign datasets:

• GTSDB
• VTSDB100

DVAL is configured as a two-stage selection pipeline in this codebase:

1. ddus builds a detector-side uncertainty shortlist from unlabeled images.
2. maple re-ranks that shortlist with crop-level MaPLe scores.

Requirements

• Python 3.9+
• CUDA-capable GPU is recommended
• uv package manager
• YOLO weights such as yolo11n.pt
• Prepared dataset folders for GTSDB and/or VTSDB100

Install dependencies:

uv sync

If your environment uses process titles, create the training environment file:

cp .env.example .env.training

At minimum, make sure PROCTITLE_STARTSTR is set before running experiments:

Datasets

The supported config folders are intentionally limited to:

configs/
+-- gtsdb/
+-- vtsdb100/

Expected dataset YAML files:

• GTSDB/data.yaml
• VTSDB100/data.yaml

Each dataset YAML should follow the YOLO format with train/val/test image paths and class names.

Run DVAL

Run DVAL on VTSDB100:

uv run python scripts/run_experiment.py --config configs/vtsdb100/config_dval.yaml

Run DVAL on GTSDB:

uv run python scripts/run_experiment.py --config configs/gtsdb/config_dval.yaml

Useful overrides:

uv run python scripts/run_experiment.py \
  --config configs/vtsdb100/config_dval.yaml \
  --device 0 \
  --seed 1

Available Configs

VTSDB100:

• configs/vtsdb100/config_dval.yaml
• configs/vtsdb100/config_random.yaml
• configs/vtsdb100/config_entropy.yaml
• configs/vtsdb100/config_coreset.yaml

GTSDB:

• configs/gtsdb/config_dval.yaml
• configs/gtsdb/config_random.yaml
• configs/gtsdb/config_entropy.yaml
• configs/gtsdb/config_coreset.yaml

DVAL Configuration

The main DVAL config uses chained strategies:

strategy: ["ddus", "maple"]
expand_ratios: [2.0]

This means each round selects a DDUS shortlist of samples_per_round * 2, then MaPLe re-ranks the shortlist down to the final samples_per_round images.

Important fields:

• initial_labeled_count: initial labeled training images.
• samples_per_round: final images selected per active-learning round.
• max_rounds: number of active-learning rounds.
• strategy_args.ddus: detector uncertainty and class/localization quality weights.
• strategy_args.maple: crop-level MaPLe scoring weights.
• maple_training: prompt-learning settings for the MaPLe checkpoint used each round.

Project Layout

.
+-- configs/
|   +-- gtsdb/
|   +-- vtsdb100/
+-- scripts/
|   +-- run_experiment.py
|   +-- train.py
|   +-- strategy.py
|   +-- setup_data.py
|   +-- simulate_labeling.py
|   +-- create_yolo_crops.py
|   +-- train_maple_round.py
+-- src/
|   +-- data/
|   +-- models/
|   +-- strategies/

Notes Before Pushing to GitHub

Avoid committing generated or heavy files such as:

• datasets: GTSDB/, VTSDB100/
• experiment outputs: full_experiments/, runs/, wandb/
• model weights: *.pt, *.pth
• local environments: .venv/, .env, .env.training

Keep only source code, configs, tests, and documentation in the repository.

License

This project is licensed under the Apache License 2.0. See LICENSE for details.