This repository provides the official implementation (partial release) of our paper:

Slide Deformable Transformer for High-Precision LiDAR Point Cloud Compression Under Review

Dynamic LiDAR point cloud compression with range images aims to reduce storage and transmission costs while preserving both spatial accuracy and temporal consistency across frames.

To tackle the limitations of existing Transformer-based methods—such as feature misalignment under motion displacement and inefficiency of global attention—we propose a new framework, SDT-PCC, which integrates:

• Slide Deformable Transformer (SDT):

  • Restricts attention to local sliding windows, capturing fine-grained correspondences.
  • Incorporates deformable convolution into cross-frame attention for adaptive motion alignment.

• Radix-Decomposition Multi-Channel Quantizer (RDMCQ):

  • Decomposes 16-bit range values into multiple channels.
  • Progressively refines precision across radix levels, mitigating quantization loss.

Experiments on the SemanticKITTI dataset demonstrate that SDT-PCC achieves higher efficiency, temporal coherence, and reconstruction accuracy compared to existing methods.

We provide an environment.yaml file for reproducibility.

conda env create -f environment.yaml -n TO

conda activate TO

Please ensure that CUDA and PyTorch versions are compatible with your local GPU setup.

For compression performance evaluation, we use our evaluation toolkit:

🔗 https://github.com/haoranli50/PCCEval

Clone the repository and install it using:

git clone https://github.com/haoranli50/PCCEval.git
cd PCCEval
python setup.py install

This will install the pcc_eval package into the current environment (TO), enabling metric evaluation during testing.

https://drive.google.com/drive/folders/1lGJkBokKTw83TysPUxjC0Ww7MbJEYrIR?usp=drive_link

checkpoints/20250620_2128/model_best.pth

Make sure the directory structure matches exactly; otherwise, the evaluation script may not find the checkpoint automatically.

You can directly evaluate the pretrained model using:

python eval.py -i ~/Dataset/data_odometry_velodyne/dataset/sequences/11/velodyne -s -f 100 -q 0.01

The available arguments in eval.py are:

• -i, --input_dir (str, required) Path to the input point cloud directory (e.g., a KITTI sequence folder containing .bin files).

• --checkpoint (str, default=None) Path to the model checkpoint. If not specified, the default path will be used.

• -s, --sample (flag) Whether to sample the test data instead of evaluating the entire sequence.

• -f, --front (int, default=0) Number of samples to take from the beginning of the sequence when sampling is enabled.

• -b, --beside (int, default=0) Number of samples to take from the end of the sequence when sampling is enabled.

• -q, --quantize_step (float, default=0.01) Quantization step size used during compression. Smaller values preserve higher precision.

• -p, --prefix (str, default=None) Prefix for saving output files.

• -c, --coder (str, default="zlib") Entropy coder used for residual encoding (e.g., zlib). Different coders may lead to different compression ratios and runtime performance.

Due to differences across compressai versions, the strategy for loading CDF buffers may vary.

If you encounter errors when loading the pretrained checkpoint (especially related to CDF tensors), please run:

python clean_ckpt.py

This script clears the stored CDFs inside the checkpoint, allowing them to be rebuilt automatically at runtime.

If you wish to train the model from scratch or fine-tune it:

python ddp_train.py

Before training, please configure dataset paths in config.py:

• kitti_train_config
• kitti_val_config

Make sure these paths correctly point to your local KITTI dataset directories.

Distributed Data Parallel (DDP) training is supported. Please ensure your multi-GPU environment is properly configured before launching training.