tlemo / IoU-Uniform-R-CNN

Codes for paper: "IoU-Uniform R-CNN: Breaking Through the Limitations of RPN"

IoU-Uniform R-CNN

This repository provides the codes for paper IoU-Uniform R-CNN

• This project is based on mmdetection framework.

Setup

Please follow official installation and getting_started guides.

Training

./tools/dist_train.sh ${CONFIG_FILE} ${GPU_NUM} [--validate] [other_optional_args] Note:

• Config files of IoU-Uniform R-CNN:

  • configs/faster_rcnn_r50_fpn_1x_IoU_reg.py
  • configs/faster_rcnn_r101_fpn_1x_IoU_reg.py
  • configs/cascade_rcnn_r50_fpn_IoU_reg_1x.py
  • configs/cascade_rcnn_r101_fpn_IoU_reg_1x.py
  • configs/pascal_voc/faster_rcnn_r50_fpn_1x_IoU_reg_voc0712.py
  • configs/pascal_voc/faster_rcnn_r50_fpn_1x_reg_separate_voc0712.py
  • configs/pascal_voc/cascade_rcnn_r50_fpn_1x_IoU_reg_voc0712.py

• We train IoU-Uniform R-CNN and accompanied detectors with 2 GPUs and 2 img/GPU. According to the Linear Scaling Rule, you need to set the learning rate proportional to the batch size if you use different GPUs or images per GPU.

Testing

./tools/dist_test.sh ${CONFIG_FILE} ${CHECKPOINT_FILE} ${GPU_NUM} [--out ${RESULT_FILE}] [--eval ${EVAL_METRICS}]

TODO

• Release IoU-Uniform R-CNN code base
• Release trained models