abrams90/soft-nms

Object Detection

Soft-NMS

This repository includes the code for Soft-NMS. Soft-NMS is integrated with two object detectors, R-FCN and Faster-RCNN. The Soft-NMS paper can be found here.

Soft-NMS+D-RFCN repository gets 40.9% mAP on COCO.

To test the models with soft-NMS, clone the project and test your models as in standard object detection pipelines. This repository supports Faster-RCNN and R-FCN where an additional flag can be used for soft-NMS.

The flags are as follows,

1. Standard NMS. Use flag TEST.SOFT_NMS 0
2. Soft-NMS with linear weighting. Use flag TEST.SOFT_NMS 1 (this is the default option)
3. Soft-NMS with Gaussian weighting. Use flag TEST.SOFT_NMS 2

In addition, you can specify the sigma parameter for Gaussian weighting and the threshold parameter for linear weighting. Detections below 0.001 are discarded. For integrating soft-NMS in your code, refer to cpu_soft_nms function in lib/nms/cpu_nms.pyx and soft_nms wrapper function in lib/fast_rcnn/nms_wrapper.py. You can also implement your own weighting function in this file.

For testing a model on COCO or PASCAL, use the following script

./tools/test_net.py --gpu ${GPU_ID} \
  --def models/${PT_DIR}/${NET}/rfcn_end2end/test_agnostic.prototxt \
  --net ${NET_FINAL} \
  --imdb ${TEST_IMDB} \
  --cfg experiments/cfgs/rfcn_end2end_ohem_${PT_DIR}.yml \
  --set TEST.SOFT_NMS 1 # performs soft-NMS with linear weighting
  ${EXTRA_ARGS}

GPU_ID is the GPU you want to test on

NET_FINAL is the caffe-model to use

PT_DIR in {pascal_voc, coco} is the dataset directory

DATASET in {pascal_voc, coco} is the dataset to use

TEST_IMDB in {voc_0712_test,coco_2014_minival,coco_2014_test} is the test imdb

TEST.SOFT_NMS in {0,1,2} is flag for different NMS algorithms. 0 is standard NMS, 1 performs soft-NMS with linear weighting and 2 performs soft-NMS with gaussian weighting

Please refer to py-R-FCN-multiGPU for details about setting up object detection pipelines. The Soft-NMS repository also contains code for training these detectors on multiple GPUs. The position sensitive ROI Pooling layer is updated so that interpolation of bins is correct, like ROIAlign in Mask RCNN. The COCO detection model for R-FCN can be found here. All other detection models used in the paper are publicly available.

Results on MS-COCO

	training data	test data	mAP@[0.5:0.95]
R-FCN, NMS	COCO 2014 train+val -minival	COCO 2015 minival	33.9%
R-FCN, Soft-NMS L	COCO 2014 train+val -minival	COCO 2015 minival	34.8%
R-FCN, Soft-NMS G	COCO 2014 train+val -minival	COCO 2015 minival	35.1%
F-RCNN, NMS	COCO 2014 train+val -minival	COCO 2015 test-dev	24.4%
F-RCNN, Soft-NMS L	COCO 2014 train+val -minival	COCO 2015 test-dev	25.5%
F-RCNN, Soft-NMS G	COCO 2014 train+val -minival	COCO 2015 test-dev	25.5%

R-FCN uses ResNet-101 as the backbone CNN architecture, while Faster-RCNN is based on VGG16.

Citing Soft-NMS

If you find this repository useful in your research, please consider citing:

@article{1704.04503,
  Author = {Navaneeth Bodla and Bharat Singh and Rama Chellappa and Larry S. Davis},
  Title = {Improving Object Detection With One Line of Code},
  Journal = {arXiv preprint arXiv:1704.04503},
  Year = {2017}
}