BiLoc: Learning 1-Bit LiDAR-based Localization with Auxiliary Objective

This repository contains the official implementation of BiLoc: Learning 1-Bit LiDAR-based Localization with Auxiliary Objective.

BiLoc is designed for always-on LiDAR localization on resource-constrained platforms. It constrains both weights and activations to 1 bit, and uses a training-only auxiliary objective to improve the binary encoder with guidance from an offline real-valued teacher. The auxiliary objective is removed at inference time, so it adds no deployment overhead.

Results

The paper evaluates BiLoc on Oxford Radar RobotCar and NCLT using mean position error and mean orientation error.

Dataset	Method	Bits	Average error
Oxford	BiLoc	1	7.56 m / 1.91 deg
NCLT	BiLoc	1	3.14 m / 4.36 deg

Qualitative Visualization

Repository Layout

This release keeps only the code needed to train and evaluate BiLoc, plus the runtime dependencies required by the model.

BiLoc/
├── biloc/                 # BiLoc training, KD training, evaluation, and deploy/ (prebuilt 1-bit GPU runtimes)
├── BHViT/                 # Minimal BHViT backbone dependency used by BiLoc
├── diffoc_full_precision/ # Full-precision DiffLoc teacher code for KD training
├── img/                   # README figures
├── LICENSE
└── README.md

The trained BiLoc checkpoints are hosted separately on Google Drive; see Model Zoo.

Environment

The experiments in the paper were run on a single NVIDIA RTX 5090 GPU.

Verified local environment:

• Python: 3.9
• PyTorch: 2.8.0+cu128
• CUDA runtime used by PyTorch: 12.8

Key Python packages in the tested environment:

Package	Version
omegaconf	2.3.0
hydra-core	1.3.2
einops	0.8.1
tensorboardX	2.6.4
timm	1.0.22
opencv-python	4.12.0
transforms3d	0.4.2
open3d	0.19.0
h5py	3.14.0
numpy	2.0.1

Activate the environment before running training or evaluation:

conda activate biloc
cd biloc

Data Preparation

BiLoc uses LiDAR data from Oxford Radar RobotCar and NCLT.

The configs expect this structure by default:

data_root/
├── Oxford/
│   ├── Oxford_pose_stats.txt
│   ├── train_split.txt
│   ├── valid_split.txt
│   └── ...
└── NCLT/
    ├── NCLT_pose_stats.txt
    ├── train_split.txt
    ├── valid_split.txt
    └── ...

The current configs use:

train:
  dataroot: ../data
  steps: 3
  skip: 2
  image_size: [32, 512]

Change train.dataroot in cfgs/oxford.yaml or cfgs/nclt.yaml to your local dataset root.

Preprocessing helpers are in:

biloc/preprocess/

Training

Run all main commands from:

cd biloc

Train the 1-bit student without the auxiliary objective:

python train.py

Train BiLoc with the auxiliary objective:

python train_kd.py

Important config fields:

ckpt: path/to/student_checkpoint.pth
exp_dir: log/your_experiment

train:
  dataset: Oxford       # Oxford or NCLT
  dataroot: path/to/data_root
  batch_size: 40
  epochs: 120

KD:
  teacher_cfg: ../diffoc_full_precision/cfgs/oxford.yaml
  teacher_ckpt: path/to/full_precision_teacher.pth
  weight: 0.8
  struct_weight: 0.05
  loss_type: entropy
  struct_loss_type: lckt

The paper uses lambda1 = 0.80 and lambda2 = 0.05 for the auxiliary objective. The code maps these to KD.weight and KD.struct_weight.

Some entry scripts currently load a config at the bottom of the file, for example:

conf = OmegaConf.load("cfgs/nclt.yaml")

Switch this line to cfgs/oxford.yaml or cfgs/nclt.yaml before launching a run.

Evaluation

Set ckpt, train.dataset, train.dataroot, and exp_dir in the selected config, then run:

python test.py

The evaluator writes:

error_t.txt
error_q.txt
pred_t.txt
gt_t.txt
pred_q.txt
gt_q.txt
trajectory.png

under cfg.exp_dir.

Model Zoo

The trained BiLoc (1-bit student) checkpoints for Oxford Radar RobotCar and NCLT are available on Google Drive. These are the checkpoints used for the BiLoc results in the paper.

Dataset	File	DDIM steps	Paper result
Oxford	biloc_oxford.pth	10	7.56 m / 1.91 deg
NCLT	biloc_nclt.pth	15	3.14 m / 4.36 deg

Place the files as follows (paths relative to the repository root):

checkpoints/
├── biloc_oxford.pth
└── biloc_nclt.pth

biloc/cfgs/oxford.yaml and biloc/cfgs/nclt.yaml already point to ../checkpoints/biloc_oxford.pth and ../checkpoints/biloc_nclt.pth.

1-Bit GPU Inference (Oxford)

biloc/deploy/ provides prebuilt GPU runtimes for the Oxford model, in which the binary layers run as real 1-bit operations (bit-packed XOR + popcount). The bit-packed weights are included. There are two variants: fused (fully fused, fastest) and eqk (one kernel per operator).

Requirements: Linux, Python 3.9, PyTorch 2.8.0+cu128 and an NVIDIA GPU of compute capability 12.0 (sm_120).

cd biloc
python -m deploy.evaluate --dataroot ../data --variant fused   # accuracy
python -m deploy.benchmark --dataroot ../data                  # latency

Acknowledgements

This code builds on DiffLoc and BHViT components used by the BiLoc model. Please follow the licenses and citation requirements of those projects when using the corresponding components.

Citation

If this repository is useful for your research, please cite:

@inproceedings{yin2026biloc,
  title={Learning 1-Bit LiDAR-based Localization with Auxiliary Objective},
  author={Yin, Kaijie and Zhang, Zhiyuan and Gao, Tian and Zhu, Wentao and Xu, Cheng-Zhong and Kong, Hui},
  booktitle={European Conference on Computer Vision},
  year={2026}
}

Contact

Any problem, please contact the first author (Email: k-yin at mi.t.u-tokyo.ac.jp)