There are five weakly supervised networks in Simple Does It: Weakly Supervised Instance and Semantic Segmentation, by Khoreva et al. (CVPR 2017). Respectively, Naive, Box, Box^i, Grabcut+, M∩G+. All of them use cheap-to-generate label, bounding box, during training and don't need other informations except image during testing.
This repo contains a TensorFlow implementation of Grabcut version of semantic segmentation.
- Operating System:
- Arch Linux 4.15.15-1
- Memory
- 64GB
- CUDA:
- CUDA V9.0.176
- CUDNN:
- CUDNN 7.0.5-2
- GPU:
- GTX 1070 8G
- Nvidia driver:
- 390.25
- Python:
- python 3.6.4
- Python package:
- tqdm, bs4, opencv-python, pydensecrf, cython...
- Tensorflow:
- tensorflow-gpu 1.5.0
- Operating System:
- Ubuntu 16.04
- Memory
- 64GB
- CUDA:
- CUDA V9.0.176
- CUDNN:
- CUDNN 7
- GPU:
- GTX 1060 6G
- Nvidia driver:
- 390.48
- Python:
- python 3.5.2
- Python package:
- tqdm, bs4, opencv-python, pydensecrf, cython...
- Tensorflow:
- tensorflow-gpu 1.6.0
./Simple_does_it/
├── Dataset
│ ├── Annotations
│ ├── CRF_masks
│ ├── CRF_pairs
│ ├── Grabcut_inst
│ ├── Grabcut_pairs
│ ├── JPEGImages
│ ├── Pred_masks
│ ├── Pred_pairs
│ ├── SegmentationClass
│ └── Segmentation_label
├── Model
│ ├── Logs
│ └── models
├── Parser_
├── Postprocess
├── Preprocess
└── Util
VOCtrainval_11-May-2012
└── VOCdevkit
└── VOC2012
├── Annotations
├── ImageSets
│ ├── Action
│ ├── Layout
│ ├── Main
│ └── Segmentation
├── JPEGImages
├── SegmentationClass
└── SegmentationObject
- Put annotations in 'Annotations'
mv {PATH}/VOCtrainval_11-May-2012/VOCdevkit/VOC2012/Annotations/* {PATH}/Simple_does_it/Dataset/Annotations/
- Put images in 'JPEGImages'
mv {PATH}/VOCtrainval_11-May-2012/VOCdevkit/VOC2012/JPEGImages/* {PATH}/Simple_does_it/Dataset/JPEGImages/
- Put Ground truth in 'SegmentationClass' for computing mIoU and IoU
mv {PATH}/VOCtrainval_11-May-2012/VOCdevkit/VOC2012/SegmentationClass/* {PATH}/Simple_does_it/Dataset/SegmentationClass/
- Pretrain model
- Move files from VOC12_CKPT to 'models'
- Run test
python ./Model/model.py --restore_target 1020 - Run train (See Training for more details)
python ./Model/model.py --is_train 1 --set_name voc_train.txt --restore_target 1020 - Performance
| set | CRF | mIoU |
|---|---|---|
| train | X | 64.93% |
| train | O | 66.90% |
| val | X | 39.03% |
| val | O | 42.54% |
- Pretrain model
- Move files from VOC12_SBD_CKPT to 'models'
- Run test
python ./Model/model.py --restore_target 538 - Run train (See Training for more details)
python ./Model/model.py --is_train 1 --set_name train.txt --restore_target 538 - Performance
| set | CRF | mIoU |
|---|---|---|
| train | X | 66.87% |
| train | O | 68.21% |
| val | X | 51.90% |
| val | O | 54.52% |
- tensorflow/models
- Put vgg_16.ckpt in 'models'
Extract annotations from 'Annotations' according to 'train.txt' or 'voc_train.txt' for VOC12 + SDB or VOC12
- For VOC12 + SBD (train set size: 10582)
- This will generate a 'train_pairs.txt' for 'grabcut.py'
python ./Dataset/make_train.py - For VOC12 (train set size: 1464)
- This will generate a 'voc_train_pairs.txt' for 'grabcut.py'
python ./Dataset/make_train.py --train_set_name voc_train.txt --train_pair_name voc_train_pairs.txt
- Result of grabcut for each bounding box will be stored at 'Grabcut_inst'
- Result of grabcut for each image will be stored at 'Segmentation_label'
- Result of grabcut for each image combing with image and bounding box will be stored at 'Grabcut_pairs'
- Note: If the instance hasn't existed at 'Grabcut_inst', grabcut.py will grabcut that image
- For VOC12 + SBD (train set size: 10582)
python ./Preprocess/grabcut.py - For VOC12 (train set size: 1464)
python ./Preprocess/grabcut.py --train_pair_name voc_train_pairs.txt
- The event file for tensorboard will be stored at 'Logs'
- Train on VOC12 + SBD (train set size: 10582)
- This will consume lot of memory.
- The train set is so large.
- Data dtyp will be casted from np.uint8 to np.float16 for mean substraction.
- Eliminate mean substraction for lower memory usage.
- Change the dtype in ./Dataset/load.py from np.float16 to np.uint8
- Comment mean substraction in ./Model/model.py
python ./Model/model.py --is_train 1 --set_name train.txt - This will consume lot of memory.
- Train on VOC12 (train set size: 1464)
python ./Model/model.py --is_train 1 --set_name voc_train.txt
- Result will be stored at 'Pred_masks'
- Result combing with image will be stored at 'Pred_pairs'
- Result after dense CRF will be stored at 'CRF_masks'
- Result after dense CRF combing with image will be stored at 'CRF_pairs'
- Test on VOC12 (val set size: 1449)
python ./Model/model.py --restore_target {num}
- Compute mIoU and IoU
python ./Dataset/mIoU.py
- Parse the command line argument
- Generating train.txt and test.txt according to 'JPEGImages'
- Not necessary
usage: divied.py [-h] [--dataset DATASET] [--img_dir_name IMG_DIR_NAME]
[--train_set_ratio TRAIN_SET_RATIO]
[--train_set_name TRAIN_SET_NAME]
[--test_set_name TEST_SET_NAME]
optional arguments:
-h, --help show this help message and exit
--dataset DATASET path to dataset (default: Util/../Parser_/../Dataset)
--img_dir_name IMG_DIR_NAME
name for image directory (default: JPEGImages)
--train_set_ratio TRAIN_SET_RATIO
ratio for training set, [0,10] (default: 7)
--train_set_name TRAIN_SET_NAME
name for training set (default: train.txt)
--test_set_name TEST_SET_NAME
name for testing set (default: val.txt)
- Extract annotations from 'Annotations' according to 'train.txt'
- Content: {image name}###{image name + num + class + .png}###{bbox ymin}###{bbox xmin}###{bbox ymax}###{bbox xmax}###{class}
- Example: 2011_003038###2011_003038_3_15.png###115###1###233###136###person
usage: make_train.py [-h] [--dataset DATASET]
[--train_set_name TRAIN_SET_NAME]
[--ann_dir_name ANN_DIR_NAME]
[--train_pair_name TRAIN_PAIR_NAME]
optional arguments:
-h, --help show this help message and exit
--dataset DATASET path to dataset (default:
Dataset/../Parser_/../Dataset)
--train_set_name TRAIN_SET_NAME
name for training set (default: train.txt)
--ann_dir_name ANN_DIR_NAME
name for annotation directory (default: Annotations)
--train_pair_name TRAIN_PAIR_NAME
name for training pair (default: train_pairs.txt)
- Grabcut a traditional computer vision method
- Input bounding box and image then generating label for training
usage: grabcut.py [-h] [--dataset DATASET] [--img_dir_name IMG_DIR_NAME]
[--train_pair_name TRAIN_PAIR_NAME]
[--grabcut_dir_name GRABCUT_DIR_NAME]
[--img_grabcuts_dir IMG_GRABCUTS_DIR]
[--pool_size POOL_SIZE] [--grabcut_iter GRABCUT_ITER]
[--label_dir_name LABEL_DIR_NAME]
optional arguments:
-h, --help show this help message and exit
--dataset DATASET path to dataset (default:
./Preprocess/../Parser_/../Dataset)
--img_dir_name IMG_DIR_NAME
name for image directory (default: JPEGImages)
--train_pair_name TRAIN_PAIR_NAME
name for training pair (default: train_pairs.txt)
--grabcut_dir_name GRABCUT_DIR_NAME
name for grabcut directory (default: Grabcut_inst)
--img_grabcuts_dir IMG_GRABCUTS_DIR
name for image with grabcuts directory (default:
Grabcut_pairs)
--pool_size POOL_SIZE
pool for multiprocess (default: 4)
--grabcut_iter GRABCUT_ITER
grabcut iteration (default: 3)
--label_dir_name LABEL_DIR_NAME
name for label directory (default: Segmentation_label)
- Deeplab-Largefov
usage: model.py [-h] [--dataset DATASET] [--set_name SET_NAME]
[--label_dir_name LABEL_DIR_NAME]
[--img_dir_name IMG_DIR_NAME] [--classes CLASSES]
[--batch_size BATCH_SIZE] [--epoch EPOCH]
[--learning_rate LEARNING_RATE] [--momentum MOMENTUM]
[--keep_prob KEEP_PROB] [--is_train IS_TRAIN]
[--save_step SAVE_STEP] [--pred_dir_name PRED_DIR_NAME]
[--pair_dir_name PAIR_DIR_NAME] [--crf_dir_name CRF_DIR_NAME]
[--crf_pair_dir_name CRF_PAIR_DIR_NAME] [--width WIDTH]
[--height HEIGHT] [--restore_target RESTORE_TARGET]
optional arguments:
-h, --help show this help message and exit
--dataset DATASET path to dataset (default:
./Model/../Parser_/../Dataset)
--set_name SET_NAME name for set (default: val.txt)
--label_dir_name LABEL_DIR_NAME
name for label directory (default: Segmentation_label)
--img_dir_name IMG_DIR_NAME
name for image directory (default: JPEGImages)
--classes CLASSES number of classes for segmentation (default: 21)
--batch_size BATCH_SIZE
batch size for training (default: 16)
--epoch EPOCH epoch for training (default: 2000)
--learning_rate LEARNING_RATE
learning rate for training (default: 0.01)
--momentum MOMENTUM momentum for optimizer (default: 0.9)
--keep_prob KEEP_PROB
probability for dropout (default: 0.5)
--is_train IS_TRAIN training or testing [1 = True / 0 = False] (default:
0)
--save_step SAVE_STEP
step for saving weight (default: 2)
--pred_dir_name PRED_DIR_NAME
name for prediction masks directory (default:
Pred_masks)
--pair_dir_name PAIR_DIR_NAME
name for pairs directory (default: Pred_pairs)
--crf_dir_name CRF_DIR_NAME
name for crf prediction masks directory (default:
CRF_masks)
--crf_pair_dir_name CRF_PAIR_DIR_NAME
name for crf pairs directory (default: CRF_pairs)
--width WIDTH width for resize (default: 513)
--height HEIGHT height for resize (default: 513)
--restore_target RESTORE_TARGET
target for restore (default: 0)
- Compute mIoU and IoU
usage: mIoU.py [-h] [--dataset DATASET] [--set_name SET_NAME]
[--GT_dir_name GT_DIR_NAME] [--Pred_dir_name PRED_DIR_NAME]
[--classes CLASSES]
optional arguments:
-h, --help show this help message and exit
--dataset DATASET path to dataset (default:
./Dataset/../Parser_/../Dataset)
--set_name SET_NAME name for set (default: val.txt)
--GT_dir_name GT_DIR_NAME
name for ground truth directory (default:
SegmentationClass)
--Pred_dir_name PRED_DIR_NAME
name for prediction directory (default: CRF_masks)
--classes CLASSES number of classes (default: 21)
- Loading data for training / testing according to train.txt / val.txt
- Save result during testing
- Map the class to grayscale value
- Map the grayscale value to RGB
- Dense CRF a machine learning method
- Refine the result
- [1] Anna Khoreva, Rodrigo Benenson, Jan Hosang, Matthias Hein, Bernt Schiele. Simple Does It: Weakly Supervised Instance and Semantic Segmentation. CVPR 2017
- [2] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, Alan L. Yuille. Semantic Image Segmentation with Deep Convolutional Nets and Fully Connected CRFs. ICLR2015
- [3] philferriere. Weakly Supervised Segmentation with Tensorflow. Implements instance segmentation as described in Simple Does It: Weakly Supervised Instance and Semantic Segmentation, by Khoreva et al. (CVPR 2017).
- [4] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, Alan L. Yuille. DeepLab: Semantic Image Segmentation with Deep Convolutional Nets, Atrous Convolution, and Fully Connected CRFs. arXive 2016
- [5] Philipp Krähenbühl, Vladlen Koltun. Efficient Inference in Fully Connected CRFs with Gaussian Edge Potentials. NIPS 2011