Self-supervised Learning for Panoptic Segmentation of Multiple Fruit Flower Species.

(Published by IEEE Robotics and Automation Letters). Published version RAL paper and preprint version arxiv paper

codeabse progress

• Preprocess raw data
• Apply data augmentation on the training data
• Train initial panoptic model using AppleA_train
• Preapre train/test unlabeled datsets for multiple run
• Generate panoptic pseudo-labels for finetuning the initial model
• Train iteratively using pseudo labels
• Evaluate the model

Requirements:

• Detectron2
• Python 3.8
• Pytorch 1.9
• CUDA 10.2
• Pycocotools 2.0

[x] Installation

1. clone this repository and go to root folder

https://github.com/siddiquemu/ssl_flower_semantic.git
cd ssl_flower_semantic

2. create environment

pip install -r det2_requirements.yml

3. This codebase is heavily based on Detectron2 and a semantic segmentation refinement method RGR. Install both and keep RGR in the root folder

[x] Data Preprocessing

1. Download the raw data from multi-species-flower. The folder structure will be

./dataseet/raw_data/
├── imgs
│   ├── AppleA
│   ├── AppleA_train
│   ├── AppleB
│   ├── Peach
│   └── Pear
└── labels
    ├── AppleA
    │   └── gt_frames
    ├── AppleA_train
    │   └── gt_frames
    ├── AppleB
    │   └── gt_frames
    ├── Peach
    │   └── gt_frames
    └── Pear
        └── gt_frames

2. run the following script from root to generate the train/test split for CV experiments

python ./dataset/data_aug_train_CV.py --CV 1 --dataset AppleA

For example the data folder structure for CV=1 in data root directory will be as follows

./dataset/ssl_data/

├── AppleA
│   └── CV1
│       ├── test_imgs
│       └── train_imgs
├── AppleB
│   └── CV1
│       ├── test_imgs
│       └── train_imgs
├── Peach
│   └── CV1
│       ├── test_imgs
│       └── train_imgs
└── Pear
    └── CV1
        ├── test_imgs
        └── train_imgs

[x] Test

1. To test the models, download CV1 models from models

├── SL
│   └── AppleA_train
├── SSL
│   ├── AppleA
│   │   └── CV1
│   │       └── iter3
│   ├── AppleB
│   │   └── CV1
│   │       └── iter3
│   ├── Peach
│   │   └── CV1
│   │       └── iter6
│   └── Pear
│       └── CV1
│           └── iter3
└── SSL_RGR
    ├── AppleA
    │   └── CV1
    │       └── iter2
    ├── AppleB
    │   └── CV1
    │       └── iter2
    ├── Peach
    │   └── CV1
    │       └── iter6
    └── Pear
        └── CV1
            └── iter3

2. run the following script to evaluate the CV models

python utils/sliding_windows_RGR.py --CV 1 --data_set AppleB --ssl_iter 3 --isLocal 1 --gpu_id 0 --model_type SSL

[x] Train

[x] To train the SL model using augmented AppleA train set:

1. run the following script from root to prepare training data

python ./dataset/data_aug_train.py --dataset AppleA_train

2. go to root directory and run

for ITER in ssl_iter; do   bash train_ssl_2gpus.sh model_type ${ITER} --label_percent GPUS data_set CV; done
for ITER in 0; do   bash train_ssl_2gpus.sh SL ${ITER} 100 2 AppleA_train 0; done

[x] To train the SSL model on the unlabeled data using AppleA trained model:

1. go to root directory and run

for ITER in 1 2 3; do   bash train_ssl_2gpus.sh SSL ${ITER} 100 2 AppleA_train 2; done

Citing ssl_flower_semantic

If you find this work helpful in your research, please cite using the following bibtex

@ARTICLE{9928359,
  author={Siddique, Abubakar and Tabb, Amy and Medeiros, Henry},
  journal={IEEE Robotics and Automation Letters}, 
  title={Self-Supervised Learning for Panoptic Segmentation of Multiple Fruit Flower Species}, 
  year={2022},
  volume={7},
  number={4},
  pages={12387-12394},
  doi={10.1109/LRA.2022.3217000}}