Awesome

MLKP

CVPR18 Paper: Multi-scale Location-aware Kernel Representation for Object Detection. Paper can be found in arXiv and CVPR2018.

MLKP is a novel compact, location-aware kernel approximation method to represent object proposals for effective object detection. Our method is among the first which exploits high-order statistics in improving performance of object detection. The significant improvement over the first-order statistics based counterparts demonstrates the effectiveness of the proposed MLKP.

Citing

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

@InProceedings{Wang_2018_CVPR,
author = {Wang, Hao and Wang, Qilong and Gao, Mingqi and Li, Peihua and Zuo, Wangmeng},
title = {Multi-Scale Location-Aware Kernel Representation for Object Detection},
booktitle = {The IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
month = {June},
year = {2018}
}

@article{wang2021multi,
  title={Multi-scale structural kernel representation for object detection},
  author={Wang, Hao and Wang, Qilong and Li, Peihua and Zuo, Wangmeng},
  journal={Pattern Recognition},
  volume={110},
  pages={107593},
  year={2021},
  publisher={Elsevier}
}

The code is modified from py-faster-rcnn.

For multi-gpu training, please refer to py-R-FCN-multiGPU

Machine configurations

OS: Linux 14.02
GPU: TiTan 1080 Ti
CUDA: version 8.0
CUDNN: version 5.0

Slight changes may not results instabilities

PASCAL VOC detection results

We have re-trained our networks and the results are refreshed as belows:

VOC07_Test set results

Networks	mAP	aero	bike	bird	boat	bottle	bus	car	cat	chair	cow	table	dog	horse	mbike	person	plant	sheep	sofa	train	tv
VGG16	78.4	80.4	83.0	77.6	70.0	71.8	84.2	87.5	86.7	67.0	83.1	70.3	84.9	85.5	81.9	79.2	52.6	79.7	79.6	81.7	81.4
ResNet	81.0	80.3	87.1	80.8	73.5	71.6	86.0	88.4	88.8	66.9	86.2	72.8	88.7	87.4	86.7	84.3	56.7	84.9	81.0	86.7	81.7

VOC12_Test set results

Networks	mAP	aero	bike	bird	boat	bottle	bus	car	cat	chair	cow	table	dog	horse	mbike	person	plant	sheep	sofa	train	tv
VGG16	75.5	86.4	83.4	78.2	60.5	57.9	80.6	79.5	91.2	56.4	81.0	58.6	91.3	84.4	84.3	83.5	56.5	77.8	67.5	83.9	67.4
ResNet	78.0	87.2	85.6	79.7	67.3	63.3	81.2	82.0	92.9	60.2	82.1	61.0	91.2	84.7	86.6	85.5	60.6	80.8	69.5	85.8	72.4

Results can be found at VGG16 and ResNet

MS COCO detection results

Networks	Avg.Precision,IOU:	Avg.Precision,Area:	Avg.Recal,#Det:	Avg.Recal,Area:
	0.5:0.95 0.50 0.75	Small Med. Large	1 10 100	Small Med. Large
VGG16	26.9 48.4 26.9	8.6 29.2 41.1	25.6 37.9 38.9	16.0 44.1 59.0
ResNet	30.0 51.3 31.0	9.6 32.4 47.2	27.8 40.7 41.7	16.4 46.8 65.1

MLKP Installation

Clone the MLKP repository

git clone https://github.com/Hwang64/MLKP.git

Build Caffe and pycaffe

cd $MLKP_ROOT
git clone https://github.com/Hwang64/caffe-mlkp.git
cd caffe-mlkp
make clean
make all -j16 && make pycaffe

Build the Cython modules
```
cd $MLKP_ROOT/lib
make
```

installation for training and testing models on PASCAL VOC dataset

3.0 The PASCAL VOC dataset has the basic structure:

 $VOCdevkit/                           # development kit
 $VOCdevkit/VOCcode/                   # VOC utility code
 $VOCdevkit/VOC2007                    # image sets, annotations, etc.

3.1 Create symlinks for the PASCAL VOC dataset

 cd $MLKP_ROOT/data
 ln -s $VOCdevkit VOCdevkit2007
 ln -s $VOCdevkit VOCdevkit2012

For more details, please refer to py-faster-rcnn.

Test with PASCAL VOC dataset

We provide PASCAL VOC 2007 pretrained models based on VGG16 and ResNet, please download the models manully from BaiduYun or GoogleDrive and put them in $MLKP_ROOT/output/

4.0 Test VOC07 using VGG16 network

python ./tools/test_net.py --gpu 0\
         --def models/VGG16/test.prototxt\
         --net output/VGG16_voc07_test.caffemodel\
         --imdb voc_2007_test\
         --cfg experiments/cfgs/faster_rcnn_end2end.yml

The final results of the model is mAP=78.4%

4.1 Test VOC07 using ResNet-101 network

python ./tools/test_net.py --gpu 0\
          --def models/ResNet/test.prototxt\
          --net output/ResNet_voc07_test.caffemodel\
          --imdb voc_2007_test\
          --cfg experiments/cfgs/faster_rcnn_end2end.yml

The final results of the model is mAP=81.0%

Train with PASCAL VOC dataset

Please download ImageNet-pretrained models first and put them into $data/ImageNet_models.

5.0 Train using single GPU

python ./tools/train_net.py --gpu 0\ 
        --solver models/VGG16/solver.prototxt\
        --weights data/ImageNet_models/VGG16.v2.caffemodel\
        --imdb voc_2007_trainval+voc_2012_trainval\ 
        --cfg experiments/cfgs/faster_rcnn_end2end.yml

5.1 Train using multi-GPUs

python ./tools/train_net_multi_gpu.py --gpu 0,1,2,3\
        --solver models/VGG16/solver.prototxt\
        --weights data/ImageNet_models/VGG16.v2.caffemodel\
        --imdb voc_2007_trainval+voc_2012_trainval\
        --cfg experiments/cfgs/faster_rcnn_end2end.yml