SFMNet: Sparse Focal Modulation for 3D Object Detection

This is the official code release of SFMNet (WACV 2026) for Waymo and Argoverse 2 datasets. The code for the nuScenes dataset can be found here.

Results on Waymo

Validation set

Method	mAP/mAPH L1	mAP/mAPH L2	Vehicle L1	Vehicle L2	Pedestrian L1	Pedestrian L2	Cyclist L1	Cyclist L2
SFMNet	81.8/79.8	75.7/73.8	80.5/80.0	72.6/72.2	85.0/80.6	77.4/73.2	79.9/78.9	77.0/76.1

Results on Argoverse 2

Validation set

Method	mAP	Vehicle	Bus	Ped.	Stop Sign	Box Truck	Bollard	C-Barrel	Motorcyclist	MPC-Sign	Motorcycle	Bicycle	A-Bus	S-Bus	Truck Cab	C-Cone	V-Trailer	Sign	Large V.	Stroller	Bicyclist	Truck	MBT	Dog	W-chair	W-Device	W-Ride
SFMNet	40.4	79.0	49.5	72.3	49.3	46.1	65.1	73.8	55.1	53.2	62.8	54.2	30.2	44.8	26.7	57.9	31.7	20.9	8.3	33.4	44.4	21.6	0.0	22.1	3.1	33.1	12.1

Installation

Requirements

The codes were tested in the following environment:

• Ubuntu 20.04
• Python 3.9
• PyTorch 1.12
• CUDA 11.3

Install pcdet

Install spconv with pip, see the official documents of spconv. Install this pcdet library and its dependent libraries by running the following:

python setup.py develop

Training & Evaluation

Change directory to the 'tools' directory:

cd tools/

• Train with a single GPU:

python train.py --cfg_file ${CONFIG_FILE}

• Train with multiple GPUs:

bash scripts/dist_train.sh ${NUM_GPUS} --cfg_file ${CONFIG_FILE}

Can add extra command line parameters. For example, this runs on 8 GPUs with a batch size of 32 for 24 epochs:

bash scripts/dist_train.sh 8 --cfg_file cfgs/sfmnet_models/sfmnet_waymo.yaml --batch_size 32 --epochs 24 --extra_tag my_run

• Evaluate a pretrained model with a single GPU:

python test.py --cfg_file ${CONFIG_FILE} --batch_size ${BATCH_SIZE} --ckpt ${CKPT}

• Evaluate with multiple GPUs:

bash scripts/dist_test.sh ${NUM_GPUS} --cfg_file ${CONFIG_FILE} --batch_size ${BATCH_SIZE}

Citation

@article{shrout2025sfmnet,
  title={SFMNet: Sparse Focal Modulation for 3D Object Detection},
  author={Shrout, Oren and Tal, Ayellet},
  journal={arXiv preprint arXiv:2503.12093},
  year={2025}
}