🚗 Diffusion-FS: Multimodal Free-Space Prediction via Diffusion for Autonomous Driving

Authors: Keshav Gupta, Tejas S. Stanley, Pranjal Paul, Arun K. Singh, K. Madava Krishna

IROS 2025

📋 Table of Contents

• 🚀 Quick Start
• 🐍 Environment Setup
• 📊 Dataset Preparation
• 🔮 Inference
• 📈 Evaluation
• 🏋️ Training
• 🤝 Contributing

🚀 Quick Start

This repository implements Diffusion-FS, a novel approach for multimodal free-space prediction in autonomous driving using diffusion models. From a dataset of raw driving logs containing image and ego trajectory pairs, our self supervised method processes such an unannotated dataset to generate freespace segments essential for autonomous driving. At inference, our model denoises a fixed number of noise samples into freespace segments. We showcase predictions across various weather conditions, times of day, road topologies, and obstacle layouts. We show the results of our method on both simulated (CARLA) and real-world (NuScenes) datasets.

🐍 Environment Setup

Prerequisites

• Python 3.10.16
• CUDA-compatible GPU (recommended)
• Conda package manager

Installation Steps

1. Create and activate conda environment:

conda create -n diff_fs_env python=3.10.16 -y
conda activate diff_fs_env

2. Install PyTorch with CUDA support:

pip install torch==2.3.1+cu118 torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118

Note: Replace cu118 with your CUDA version (e.g., cu121 for CUDA 12.1)

3. Install additional dependencies:

pip install -r requirements.txt

📊 Dataset Preparation

We perform experiments on both CARLA (simulated) and NuScenes (real-world) datasets. Note that while data generation, we use the given 3D bounding boxes for freespace creation but it is equivalent to use the 2D bounding boxes instead. Notably, the following 2 things are equivalent 1) The freespace can be limited to the nearest obstacle after projecting the unrestricted freespace segment created using the ego-trajectory in the image space and then limiting it to the closest 2D obsacle bounding box. 2) The freespace can be first limited to the nearest obstacle using the 3D bounding boxes or locations of the obstacles in the BEV, and then projecting this to the image space.

🏎️ CARLA Dataset

Option 1: Download Pre-cached Dataset (Recommended)

1. Download from HuggingFace Dataset Repository
2. Contains cached training/evaluation data and scenario-wise situation classes (no lane, single lane, multilane, intersection)
3. Merge the splits:

mv carla/carla_train/split_*/*.npz carla/carla_train/
rm -r carla/carla_train/split_*

Option 2: Generate from LAV Dataset

1. Download data from LAV repository
2. Cache manually using datasets_all/fs_dataset.py
3. Update data_dir_train and data_dir_test in datasets_all/fs_cached.yaml

🌍 NuScenes Dataset

Option 1: Download Pre-cached Dataset (Recommended)

1. Download from HuggingFace Dataset Repository
2. Merge the splits:

mv nuscenes/FS/split_*/*.npz nuscenes/FS/
mv nuscenes/fs_meta/split_*/*.npz nuscenes/fs_meta/
rm -r nuscenes/FS/split_*
rm -r nuscenes/fs_meta/split_*

Option 2: Generate Yourself

Use the script: dataset_all/nusc_save.py

🔮 Inference

📥 Download Pretrained Models

Download checkpoints from Google Drive:

• CARLA Base: pretrained_ckpts/carla_base.ckpt
• CARLA Command Conditioning: pretrained_ckpts/carla_cls.ckpt
• NuScenes Base: pretrained_ckpts/nuscenes_base.ckpt

🎯 Run Inference

python3 infer.py

Configuration Options (modify at the beginning of infer.py):

• Config file path
• Checkpoint path
• Output visualization path
• Noise template usage (optional)
• Obstacle guidance during inference (optional)

Output: 6 inferred freespace segments overlayed on each input image.

📈 Evaluation

🏎️ CARLA Evaluation

Basic Evaluation

python3 evaluate.py

Directional Deviation Metrics

python3 dd_calc.py

Comprehensive Evaluation (All 5 Experiments)

Run all experiments (base, obs, cls, noise, obs+cls) sequentially:

python3 eval_all_dd.py

Parse and display results:

python3 calc.py

Configuration Options (modify at the beginning of each script):

• Checkpoint path
• Config file path
• Noise template usage (optional)
• Obstacle guidance during inference (optional)
• Situation classes folder path (for eval_all_dd.py)

Precomputed Results: Available in results/ folder

🌍 NuScenes Evaluation

python3 evaluate.py

Use the same optional arguments as CARLA evaluation.

🏋️ Training

Configuration

Modify hyperparameters in:

• configs/carla.yaml (for CARLA dataset)
• configs/nuscenes.yaml (for NuScenes dataset)

Available Parameters:

• Model hyperparameters
• Optimizer settings
• Checkpointing paths
• Diffusion denoising parameters

Start Training

python3 train.py

🤝 Contributing

We welcome contributions! Please feel free to submit issues and pull requests.

📄 Citation

If you find this work useful, please cite our paper:

@misc{gupta2025diffusionfsmultimodalfreespaceprediction,
      title={Diffusion-FS: Multimodal Free-Space Prediction via Diffusion for Autonomous Driving}, 
      author={Keshav Gupta and Tejas S. Stanley and Pranjal Paul and Arun K. Singh and K. Madhava Krishna},
      year={2025},
      eprint={2507.18763},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2507.18763}, 
}

---

📧 Contact: For questions or issues, please open an issue on GitHub or contact the authors.

🌟 Star this repo if you find it helpful!