Affordance detection with pytorch

this repository is the implementation to detect affordances of objects with pytorch

requirements

• python 3
• pytorch >= 0.4
• tensorboardx
• scipy, skimage, tqdm

dataset

use the following dataset:

Part Affordance Dataset

Affordance Detection of Tool Parts from Geometric Features,
Austin Myers, Ching L. Teo, Cornelia Fermüller, Yiannis Aloimonos.
International Conference on Robotics and Automation (ICRA). 2015.

training on Part Affordance Dataset

please run train.py in the command line after downloading the dataset in 'part-affordance-dataset' directory.

python train.py -h

Then,

usage: train.py [-h] [--model MODEL] [--class_weight CLASS_WEIGHT]
                [--batch_size BATCH_SIZE] [--num_worker NUM_WORKER]
                [--max_epoch MAX_EPOCH] [--learning_rate LEARNING_RATE]
                [--in_channel IN_CHANNEL] [--n_classes N_CLASSES]
                [--device DEVICE] [--writer WRITER]
                [--result_path RESULT_PATH]

train network for affordance detection

optional arguments:
  -h, --help            show this help message and exit
  --model MODEL         available model options => FCN8s/SegNetBasic/UNet
  --class_weight CLASS_WEIGHT
                        if you want to use class weight, input True. Else,
                        input False
  --batch_size BATCH_SIZE
                        number of batch size: number of samples sent to the
                        network at a time
  --num_worker NUM_WORKER
                        number of workers for multithread data loading
  --max_epoch MAX_EPOCH
                        the number of epochs for training
  --learning_rate LEARNING_RATE
                        base learning rate for training
  --in_channel IN_CHANNEL
                        the number of the channel of input images
  --n_classes N_CLASSES
                        number of classes in the dataset including background
  --device DEVICE       the device you'll use (cpu or cuda:0 or so on)
  --writer WRITER       if you want to use SummaryWriter in tesorboardx, input
                        True. Else, input False
  --result_path RESULT_PATH
                        select your directory to save the result

you can choose the model, batch size and so on as you like

predict

To predict affordance of images in the dataset, run python predict.py in the command line.
Just like trainig, you can select the parameter, path and so on.

python predict.py -h

usage: predict.py [-h] [--model MODEL] [--params_path PARAMS_PATH]
                  [--num_images NUM_IMAGES] [--in_channel IN_CHANNEL]
                  [--n_classes N_CLASSES] [--device DEVICE]
                  [--result_path RESULT_PATH]

train network for affordance detection

optional arguments:
  -h, --help            show this help message and exit
  --model MODEL         available model options => FCN8s/SegNetBasic/UNet
  --params_path PARAMS_PATH
                        if you want to use a trained model, input the path of
                        a file of it
  --num_images NUM_IMAGES
                        number of images to predict for segmentation
  --in_channel IN_CHANNEL
                        the number of the channel of input images
  --n_classes N_CLASSES
                        number of classes in the dataset including background
  --device DEVICE       the device you'll use (cpu or cuda:0 or so on)
  --result_path RESULT_PATH
                        select your directory to save the result

date

Oct. 26, 2018