This repository is the official implementation of "Task-Oriented Token Pruning for Efficient Object Detection and Segmentation" published in IROS 2025.
To install the required dependencies, run the following command:
# modify the prefix in environment.yaml as needed
conda env create -f environment.yaml
python -m pip install -e .Please refer to the mmdet repository for instructions on how to prepare datasets and pre-trained models: mmdet.
Training:
./tools/dist_train.sh configs/prune/topk/dynamic-mask-rcnn_vit-t-p4-w7_fpn_ms_1x_coco_gtguide_subset_topk-final.py 8 --auto-scale-lrEvaluation:
./tools/dist_test.sh configs/prune/topk/dynamic-mask-rcnn_vit-t-p4-w7_fpn_ms_1x_coco_gtguide_subset_topk-final.py your_checkpoint.pth 8This code is built upon the mmdet and SViT repositories. We thank the authors for their excellent work.
If you find this work useful in your research, please consider citing: