Test implementation of USD-Seg.

Contents

Requirements

Features

Preparation

Training

Evaluation

Benchmarking the speed of network

• python 3.6
• mxnet 1.5.0
• scikit-lean 0.21.3

Other common package version is not very crucial. If you use different versions of scikit-learn, you may have to retrain the dictionary by your own.

• Include both USD-Seg(Yolov3-darknet53). USD-Seg(Yolov3-tiny) is still in development but we have a positive altitude.

• Good performance

  416x416	VOC2012 Test(mAP)	SBD Test(mAP)	Time per forward (batch size = 1)
  ESE-Seg(Darknet53)	69.3%	64.1%	26 ms
  ESE-Seg(Darknet-tiny)	53.2%	48.1%	8.0 ms
  USD-Seg
  USD-Seg(tiny)

  The models are trained from bbox-pretrained weights on coco and then trained in SBD.

• Some efficient backbones on hand

  Like yolov3-tiny, yolov3.

  Check folder /gluon-cv/gluoncv/model_zoo/yolo/darknet.py for details.

git clone https://github.com/WenqiangX/ese_seg.git

git clone https://github.com/dmlc/gluon-cv.git

cd gluon-cv

python setup.py develop --user

cd ..

cp -r ./ese_seg/gluon-cv gluon-cv

cd ese_seg

• SBD (Semantic Boundaries Dataset and Benchmark)

mkdir data

cd data

Firstly, you can download our preprocessed dataset on google drive.

tar -zxvf VOCsbdche.tar.gz

• Custom Dataset

Our dataset format is equal to pascal voc 2012, you can check it in our preprocessed dataset.

JPEGImages( image data )

ImageSets ( dataname list )

cheby_fit ( cheby_coefficient train label )

label_polygon_360_xml ( polygon val label )

We provide preprocessing code in ./labelutils, you can just run your.bash simply according to sbd_dataset.bash For example, in SBD, preprocess label into (SegementationObject/ SegmentationClass/) as PascalVOC format

./sbd
    ./SegementationObject
    ./SegementationClass

make the ./labelutils and ./sbd in the same directory, such ./ESE-SEG/sbd

bash sbd_dataset.bash

Another important thing is that there will be very few invalid labels, so you can change your train/val labelname list according to ./coef_8_success.txt. We have provided the sbd vaild train and val txt.

To Do: We will release two pretrained weight based on darknet53 and tinydarknet in some weeks.

1.1) run

cd ese_seg

python sbd_train_che_8.py --syncbn --network darknet53 --batch-size 20 --dataset voc --gpus 0 --warmup-epochs 10 --save-prefix ./darknet53_result

or

python sbd_train_che_8.py --syncbn --network tiny_darknet --batch-size 64 --dataset voc --gpus 0 --warmup-epochs 10 --save-prefix ./darknet53_result

you can also change the val dataset to voc2012 by --val_2012.

The result are both very well by direct training. Your can pretrain only bbox in coco, make its bbox prediction more accurate.

1.1) Preprocess the coco dataset as custom dataset. And coco dataset is train dataset, voc/sbd is val dataset. You can check the path in ./gluon-cv/gluoncv/data/pascal_voc/detection.py

1.2) run

python sbd_train_che_8.py --syncbn --network darknet53 --only_bbox True --batch-size 20 --dataset cocopretrain --gpus 0 --warmup-epochs 10 --save-prefix ./darknet53_pretrain_coco_result

or

python sbd_train_che_8.py --syncbn --network tiny_darknet --only_bbox True --batch-size 20 --dataset cocopretrain --gpus 0 --warmup-epochs 10 --save-prefix ./darknet53_pretrain_coco_result

To Do:

cd ese_seg

1.1) Your should have a model weight $weight_path$

1.2) run

python sbd_eval_che_8.py --network darknet53 --val_voc2012 True --resume weight_path --save-prefix ./eval_voc2012_tinydarknet

or

python sbd_eval_che_8.py --network darknet53 --resume weight_path --save-prefix ./eval_voc2012_tinydarknet

2.1) Your should have a model weight $weight_path$

2.2) run

python sbd_eval_che_8.py --network tiny_darknet --val_voc2012 True --resume weight_path --save-prefix ./eval_voc2012_tinydarknet

or

python sbd_eval_che_8.py --network tiny_darknet --resume weight_path --save-prefix ./eval_voc2012_tinydarknet

cd ese-seg

1.1) Your should have a model weight $weight_path$ and $image_dir$

1.2) run

python demo_yolo.py --network yolo3_darknet53_voc --images $image_dir$ --save_dir ./demo_darknet53 --pretrained $weight_path$

2.1) Your should have a model weight $weight_path$ and $image_dir$

2.2) run

python demo_yolo.py --network yolo3_darknet53_voc --images $image_dir$ --save_dir ./demo_darknet53 --pretrained $weight_path$

We test the speed based on forward process and yolo haed predicted.

Tiny-darknet: backbone ~2ms two-head ~3+3 ms

If you find our work is useful to your research, feel free to cite

@inproceedings{xu2019ese,
  title={Explicit Shape Encoding for Real-Time Instance Segmentation},
  author={Xu, Wenqiang and Wang, Haiyang and Qi, Fubo and Lu, Cewu},
  booktitle={ICCV},
  year={2019}
}

I got a lot of code from gluon-cv, thanks to Gluoncv.

YOLOv3 in GluonCV is the best reimplemented YOLO so far, hence we build this system upon it. However, GluonCV is more a production code than research code, so even all we have to do is changing dataloader and network prediction head, it still took a lot of modification from original GluonCV. For research use, it is encouraged to use more light-weighted or more flexible detection framework, should make the implementation of our core idea easier.

ESE-SEG