CURVE

This repository contains a compact training and evaluation implementation for CURVE: Learning Causality-Inspired Invariant Representations for Robust Scene Understanding via Uncertainty-Guided Regularization.

Overview

The implementation covers the core components of CURVE, including:

• Object-level feature construction
• Variational scene graph generation
• Uncertainty-aware relational intervention
• Differentiable structure learning
• Temporal aggregation and prediction

The code supports the experimental regimes reported in the paper, including in-distribution evaluation, zero-shot out-of-distribution generalization, and sim-to-real transfer under varying supervision levels.

Installation

Use a Python environment with:

• PyTorch
• PyTorch Geometric
• scikit-learn
• pandas
• matplotlib
• wandb
• pytorch-nlp

You can install the minimal Python dependencies with:

pip install -r requirements.txt

Training

Run training and evaluation with:

python scripts/train.py --yaml_path config/main_config.yaml --input_path /path/to/dataset.pkl

To force CPU execution:

python scripts/train.py --yaml_path config/main_config.yaml --device cpu --input_path /path/to/dataset.pkl

To evaluate a saved checkpoint:

python scripts/train.py \
  --yaml_path config/main_config.yaml \
  --input_path /path/to/dataset.pkl \
  --model_load_path /path/to/model.pt

If your dataset pickle depends on classes from another repository, expose that code with:

export CURVE_EXTRA_PYTHONPATH=/path/to/external/repo

Transfer Learning

The transfer learning entrypoint remains available at scripts/transfer.py and uses config/transfer_config.yaml.

Datasets

This codebase follows the RS2G pipeline and uses RoadScene2Vec-style preprocessing for scene graph construction.

• The source-domain training setup follows the same processing protocol as RS2G.
• DeepAccident is used for zero-shot out-of-distribution evaluation.
• DoTA is used for sim-to-real transfer evaluation.

The processed datasets are not included in this repository. DeepAccident and DoTA cannot be redistributed here because we do not have permission to open-source them.