🏆 This paper "Butter: Frequency Consistency and Hierarchical Fusion for Autonomous Driving Object Detection" has been accepted to ACM Multimedia 2025.
Paper: https://www.arxiv.org/pdf/2507.13373
Checkpoint: https://huggingface.co/Christopher-Lim/Butter
Butter is a novel 2D object detection framework designed to enhance hierarchical feature representations for improved detection robustness.
Install
Install ultralytics. Note: Python >= 3.8, PyTorch >= 1.8。pip install ultralyticsUsage
KITTI Dataset
python Butter-KITTI-train.py --model_path ultralytics/cfg/models/Butter/Butter[SCALE]-All-KITTI.yaml --device 0 --name [NAME]The --model_path option specifies the specific model configuration, with predefined configurations stored in the ultralytics/cfg/models/Butter/ directory. [SCALE] You may choose between n,s,m,l,x; if [SCALE] is not specified, it defaults to n. In the original paper, Butter is a model scaled with m. The --device option specifies the GPU ID for single-GPU training. The --name option sets the name of the training experiment as [NAME], and the final results will be saved under runs/detect/[NAME].
After training is completed, the best model weights will be saved at runs/detect/[NAME]/weights/best.pt.
Cityscapes Dataset
TORCH_DISTRIBUTED_DEBUG=DETAIL python -m torch.distributed.run --nproc_per_node=4 Butter-City-train.py --model_path ultralytics/cfg/models/Butter/Butterm-All-City.yaml --device 1,2,3,6 --name [NAME]The --nproc_per_node option specifies the number of GPUs to use. The --model_path option specifies the specific model configuration, with predefined configurations stored in the ultralytics/cfg/models/Butter/ directory. The --device option specifies the GPU IDs for multi-GPU training, separated by commas. The --name option sets the name of the training experiment as [NAME], and the final results will be saved under runs/detect/[NAME].
BDD100K Dataset
TORCH_DISTRIBUTED_DEBUG=DETAIL python -m torch.distributed.run --nproc_per_node=4 Butter-BDD100K-train.py --model_path ultralytics/cfg/models/Butter/Butterm-All-BDD100K.yaml --device 4,5,6,7 --name [NAME]The --nproc_per_node option specifies the number of GPUs to use. The --model_path option specifies the specific model configuration, with predefined configurations stored in the ultralytics/cfg/models/Butter/ directory. The --device option specifies the GPU IDs for multi-GPU training, separated by commas. The --name option sets the name of the training experiment as [NAME], and the final results will be saved in runs/detect/[NAME].
COCO Dataset
TORCH_DISTRIBUTED_DEBUG=DETAIL python -m torch.distributed.run --nproc_per_node=4 Butter-coco-train.py --model_path ultralytics/cfg/models/Butter/Butter-All-COCO.yaml --device 1,2,3,6 --name [NAME]The --nproc_per_node option specifies the number of GPUs to use. The --model_path option specifies the specific model configuration, with predefined configurations stored in the ultralytics/cfg/models/Butter/ directory. The --device option specifies the GPU IDs for multi-GPU training, separated by commas. The --name option sets the name of the training experiment as [NAME], and the final results will be saved under runs/detect/[NAME].
KITTI Dataset
python Butter-KITTI-predict.py --model_path runs/detect/[TRAIN_NAME]/weights/best.pt --name [NAME]Here, the --model_path option specifies the best model weights obtained from training, located at runs/detect/[TRAIN_NAME]/weights/best.pt, where [TRAIN_NAME] should be replaced with the name of the training experiment. The --name option sets the name of the prediction experiment as [NAME], and the final results will be saved in runs/detect/[NAME].
Cityscapes Dataset
python Butter-City-predict.py --model_path runs/detect/[TRAIN_NAME]/weights/best.pt --name [NAME]Here, the --model_path option specifies the best model weights obtained from training, located at runs/detect/[TRAIN_NAME]/weights/best.pt, where [TRAIN_NAME] should be replaced with the name of the training experiment. The --name option sets the name of the prediction experiment as [NAME], and the final results will be saved in runs/detect/[NAME].
BDD100K Dataset
python Butter-BDD100K-predict.py --model_path runs/detect/[TRAIN_NAME]/weights/best.pt --name [NAME]Here, the --model_path option specifies the best model weights obtained from training, located at runs/detect/[TRAIN_NAME]/weights/best.pt, where [TRAIN_NAME] should be replaced with the name of the training experiment. The --name option sets the name of the prediction experiment as [NAME], and the final results will be saved in runs/detect/[NAME].
COCO Dataset
python Butter-COCO-predict.py --model_path runs/detect/[TRAIN_NAME]/weights/best.pt --name [NAME]Here, the --model_path option specifies the best model weights obtained from training, located at runs/detect/[TRAIN_NAME]/weights/best.pt, where [TRAIN_NAME] should be replaced with the name of the training experiment. The --name option sets the name of the prediction experiment as [NAME], and the final results will be saved in runs/detect/[NAME].