This is the official PyTorch implementation of TrafficLoc [ICCV 2025]
[Project page], [Paper]
We use Python 3.9, Pytorch 1.12.1 and CUDA 11.3.1
# clone repository
git clone https://github.com/TUM-Luk/TrafficLoc.git
# create conda environment
conda create -n trafficloc python=3.9
# install pytorch
conda install pytorch==1.12.1 torchvision==0.13.1 torchaudio==0.12.1 cudatoolkit=11.3 -c pytorch
# install required packages
pip install -r requirement.txt
We provide our Carla Intersection Dataset in Baidu Cloud Dist Link (code: g9fk). Please remain ~400GB storage space before downloading it.
For introduction of Carla Intersection Dataset, please refer to the README.md in DATASET folder.
To generate training and query files, you should following the codes below.
When generating query file, remember to run the generate_query_file.py twice. By setting the variable is_seq5 to True and False, you can generate the query file for Test split T1-T7 and T1-T7_hard
# generate the dataset folder
python DATASET/transform_dataset.py
# generate training label files
python DATASET/generate_label_file.py
python DATASET/generate_label_txt.py
# generate evaluation label files
python DATASET/generate_query_file.py # set is_seq5=False
python DATASET/generate_query_file.py # set is_seq5=True
python DATASET/generate_query_txt.py
After running the code, you should see the train_list folder in your dataset folder, which contains training and query txt. During the training, the validation set is a combination of all three Test splits.
demo_dataset
├── t1_int1
│ ├── mapping
│ ├── query
│ ├── train_list_v50_s25_io03_vo025 # new folder
│ ├── query_t1_int1_v50_s25_io03_vo025.npy
│ ├── seq5_query_t1_int1_v50_s25_io03_vo025.npy
│ ├── train_all_50_0.npy
│ ├── train_all_50_1.npy
│ ├── ......
│ ├── train_all_50_8.npy
├── t1_int2
├── t1_int3
├── ......
├── train_list # new folder
│ ├── train_allscene_v50_s25_io03_vo025.txt # Train
│ ├── query_all_3testset_v50_s25_io03_vo025.txt # Val
│ ├── query_all_1to7_v50_s25_io03_vo025.txt # Test T1-T7
│ ├── query_all_1to7_v50_s25_io03_vo025_seq5.txt # Test T1-T7_hard
│ ├── query_t10_int1_v50_s25_io03_vo025.txt # Test T10
│ ├── xxxxxx.txt # train & val file for *one* intersection
We use the prepared KITTI dataset provided by CorrI2P. Unzip the downloaded files, and the directory is as follows:
kitti_data
├── calib
├── 00
├── 01
├── ......
├── 10
├── sequences
├── 00
├── 01
├── ......
├── 10
We use the prepared Nuscenes dataset provided by CorrI2P. You can also use the script for preparing Nuscenes dataset in nuScenes_script folder from the CorrI2P repository.
nuscenes_data
├── train
├── img
├── K
├── PC
├── test
├── img
├── K
├── PC
For training on Carla Intersection Dataset
# Train on Carla Intersection Dataset
sh scripts/train_scripts/train_carla.sh
# Train on KITTI Dataset
sh scripts/train_scripts/train_kitti.sh
# Train on Nuscenes Dataset
sh scripts/train_scripts/train_nuscenes.sh
For evaluation on Carla Intersection Dataset
# Test on Carla Intersection Dataset
sh scripts/test_scripts/test_carla.sh
# Test on KITTI Dataset
sh scripts/test_scripts/test_kitti.sh
# Test on Nuscenes Dataset
sh scripts/test_scripts/test_nuscenes.sh
For pretrained DUSt3R Encoder weights, we use DUSt3R_ViTLarge_BaseDecoder_512_dpt.pth. You can download it from the DUSt3R.
For evaluation, we also provide the pretrained model.
| Modelname | Trained Dataset |
|---|---|
trafficloc_carla.pth |
Carla Intersection Dataset |
trafficloc_kitti.pth |
KITTI Odometry |
trafficloc_nuscenes.pth |
Nuscenes |
@string{iccv="IEEE International Conference on Computer Vision (ICCV)"}
@inproceedings{xia2025trafficloc,
title = {TrafficLoc: Localizing Traffic Surveillance Cameras in 3D Scenes},
author = {Y Xia and Y Lu and R Song and O Dhaouadi and JF Henriques and D Cremers},
booktitle = {IEEE International Conference on Computer Vision (ICCV)},
year = {2025},
titleurl = {trafficloc.png},
keywords = {3D Localization, Traffic Surveillance Camera, LiDAR point cloud},
url = {https://tum-luk.github.io/projects/trafficloc/},
}
We sincerely thank the authors of DeepI2P, CorrI2P, CoFiI2P for their public codes as well as their prepared KITTI and NuScenes Dataset.
We also want to thank the authors of NeuMap, as our project is built based on their public codes.