This is the official implementation of LiFT (DASIP 2025). LiFT is a simple, lightweight (20.73 GMAC) and fully-sparse 3D object detector. It is tailored for (but not restricted to) real-time implementation on low-end FPGAs. By the time of publication, it achieves the best detection performance on NuScenes-val among methods with comparable structure and complexity. For more details, please refer to:
LiFT: Lightweight, FPGA-tailored 3D object detection based on LiDAR data [Paper]
Konrad Lis, Tomasz Kryjak, Marek Gorgoń
The
Refer to the original OpenPCDet readme.
The additional requirement in LiFT is the Brevitas, which can be installed with pip install brevitas.
The LiFT is trained and validated in the same way as all the other models in the OpenPCDet repo. Refer to the original OpenPCDet readme.
| nuScenes Detection | Set | mAP | NDS | Download |
|---|---|---|---|---|
| LiFT | val | 51.84 | 61.01 | ckpt |
- The implementation on FPGA is not ready yet, but it's in a progress. Updates regarding its release will be posted in this repository.
- The LiFT is quantised using the Brevitas library, as we regard it's more flexible than native PyTorch and spconv quantisation and provides more compatible interface to FINN for future implementation on FPGA.
This project is released under the Apache 2.0 license.