RGANet

Official Pytorch implementation for the journal article:

    @article{mo2022rganet,  
        title = {Realtime Global Attention Network for Semantic Segmentation},
        author = {Mo, Xi and Chen, Xiangyu},
        journal = {IEEE Robotics and Automation Letters with ICRA 2022 Presentation},
        year = {2022},
        month = jan,
        volumn = {7},
        number = {2},
        pages = {1574-1580},
        publisher = {IEEE},
        doi = {10.1109/LRA.2022.3140443}
    }

Requirements

python >= 3.5
pytorch >= 1.0.0
(optional) thop, apex, tqdm

Demo

• Create the folder checkpoint in the root directory, download our pretrained checkpoint (53.1MB) to the folder, then run the demo:

python demo.py

Suction area predictions should be saved in the folder sample, or specify the path/to/checkpoint and path/to/samples using -c and -d args respectively.

Training

• Prepare dataset and train from scratch

Please consult utils/configuraion.py if you want to customize training setup, then download suction-based-grasping-dataset.zip (1.6GB), create the folder dataset in the root directory. There are two ways to train RGANet-NB:

Extract the main folder suction-based-grasping-dataset to dataset, run

(default) > python RGANet.py -train

Extract the main folder to somewhere else and specify the paths:

(customized) > python RGANet.py -train -i path/to/color-input -l path/to/label

• Restore training from checkpoint

By default, RGANet read the latest checkpoint from the folder checkpoint, you can also specify the checkpoint using arg -c:

python RGANet.py -train -r -c path/to/checkpoint

Test

The checkpoint is required before any test, and we only present the RGANet-NB architecture. Please consult utils/configuraion.py if you want to customize the testing, then run

python RGANet.py -test

Refer to arg -d to specify the path to save predictions.

Validation

We provide both online validation and offline valuation. The online validation runs tests on all checkpoints and estimates the checkpoint that may has the best performance. Offline validation requires predictions saved to disk beforehand, i.e., run test w/ predictions written to disk before any offline validation. Please make sure you've set desired options in utils/configuraion.py. Run
python RGANet.py -v or python RGANet.py -test -v
Refer to arg -d to specify the path to predictions.

Scripts

• In addition to the functions provided above, we also provide useful tools:

name	illustration
calculator.py	evaluation statistics, calculate model parameters
eval_adaptor.py	split items of validation result to separate files
models.py	standalone training, validation of segmentation models
pred_transform.py	convert other predictions to processable images
proportion.py	compute adaptive weights for CE loss and focal loss
seg_models.py	standalone runtime test for segmentation models

License

Apache 2.0