Skip to content

Latest commit

 

History

History
146 lines (118 loc) · 6.73 KB

readme.md

File metadata and controls

146 lines (118 loc) · 6.73 KB
Raw

The main tasks that we do here are:

  1. Training the segmentation model via all the labeled data (this will give us the upper bound performance)
  2. Computing the performance of segmentation network when we train it via the data that are labeled from the active learning process.

We use Camvid and Cityscapes dataset. Our model support FCN and DeeplabV3 backbones.

Datasets:

CamVid

Download the CamVid Dataset and put it inside the data folder as follow:

Dataset Format 
|-- /Data/camvid/
           |-- train
                |--0001TP_006690.png
                ...
           |-- trainannot
                |--0001TP_006690.png
                ...
           |-- test
                |--0001TP_008550.png
                ...
           |-- testannot
                |--0001TP_008550.png
                ...
           |-- val
                |--0016E5_07959.png
                ...
           |-- valannot
                |--0016E5_07959.png
                ...

If you have the train and trainannot folder from the active learning process for camvid, copy (or move) those folders from Data/camvid/, so it looks as follows:

|-- Data/camvid/
           |-- train_al
                ...
           |-- trainannot_al
                ...

Cityscapes

Download the Cityscapes Dataset and put it inside the data folder as follow:

Dataset Format 
           
|-- /Data/cityscapes/
           |-- leftImg8bit
                |--train
                    |--achen
                        |--aachen_000000_000019_leftImg8bit.png
                        ...
                |--test
                    ...
                |--val
                    ...
            |--gtFine
                |--train
                    |--achen
                        |--aachen_000000_000019_gtFine_labelIds.png
                        ...
                |--test
                    ...
                |--val
                    ...

Training

The scrips that we use for training our model is provided in FinalTraining/scrips/ depending on the experiment you can copy and move the sctip to FinalTraining/ and then run it in the terminal by sh nameofscript.sh. 1.4 to 1.5 we observe that the results slightly changed. The results that are provided here are slightly better than the numbers that provided on the paper.

Understandning the script

here is an example to understand the elements of script, you can also read the args for more detail.

Example of running command 
DIRECTORY=`dirname $0`
OMP_NUM_THREADS=1 python $DIRECTORY/run.py -m full \
--save_dir $DIRECTORY/SAVE_city_FCN+/  \ 
--dataset_dir  $DIRECTORY/../../cityscapes \ # provide the directory to the datase, e.g. /Data/citiscapes/
--backbone fcn \ # choose the backbone, fcn or deeplab
--dataset cityscapes  \ #select the dataset, cityscapes or camvid
--name model_city \ # choose a name for your model
--optimizer adam   \
--learning_rate 5e-4  --lr_decay 0.1    \
--weight_decay 2e-4  \
--height 512  --width 1024 \ # resize the image to the desire size
--epochs 60  \
--lr_decay_epochs 20 \
--save_val_every_epoch 1 \
--batch_size 4 \
--consistency 1 \ #enable the self-consistency
--gpunum  1 \ # gpu id that you use
--print_step

If you are using the self-consistency for cityscapes with resolusion of 512x1024 you need to have a GPU with memory of 24G. Please note that using a smaller GPU will lead to lower performance due to the limitation for batch size.

If you want to train the model for the data that you got via active learning, you need to go to EquAL/FinalTraining/dataloaders/camvid/ or EquAL/FinalTraining/dataloaders/cityscapes/ and change the train_folder and train_lbl_folder to the name that you chose for the name of your model.

Also please note that in EquAL/FinalTraining/run.py and EquAL/ActiveLearning/EquAL.py we fixed the seed for reproducibility. You can change the seed or comment it, but it is suggested to have the same seed for comparison.

Results
Dataset/Model Input Resolution Classes Batch Size Epochs Mean IoU (%) Budget Self-Consistency
Camvid (Fully Trained)-FCN 360x480 11 5 60 0.664 100% 0
Cityscapes (Fully Trained)-FCN 512x1024 19 4 60 0.643 100% 0
Camvid (Fully Trained)-FCN+ 360x480 11 5 60 0.679 100% 1
Cityscapes (Fully Trained)-FCN+ 512x1024 19 4 60 0.677 100% 1
Camvid (Active Learning)-FCN 360x480 11 5 60 0.634 12% 0
Cityscapes (Active Learning)-FCN 512x1024 19 4 60 0.622 12% 0
Camvid (Active Learning)-FCN+ 360x480 11 5 60 0.644 12% 1
Cityscapes (Active Learning)-FCN+ 512x1024 19 4 60 0.649 12% 1
Dataset/Model Input Resolution Classes Batch Size Epochs Mean IoU (%) Budget Self-Consistency
Camvid (Fully Trained)-DeepLabV3 360x480 11 5 60 0.667 100% 0
Cityscapes (Fully Trained)-DeepLabV3 512x1024 19 4 60 0.649 100% 0
Camvid (Fully Trained)-DeepLabV3+ 360x480 11 5 60 0.672 12% 1
Cityscapes (Fully Trained)-DeepLabV3+ 512x1024 19 4 60 0.697 12% 1
Camvid (Fully Trained)-DeepLabV3 360x480 11 5 60 0.622 12% 0
Cityscapes (Fully Trained)-DeepLabV3 512x1024 19 4 60 12% 0
Camvid (Fully Trained)-DeepLabV3+ 360x480 11 5 60 0.634 12% 1
Cityscapes (Fully Trained)-DeepLabV3+ 512x1024 19 4 60 12% 1