DOS Dataset

DOS Dataset is a novel indoor deformable object segmentation dataset for sweeping robots, introduced in the paper "DOS Dataset: A Novel Indoor Deformable Object Segmentation Dataset for Sweeping Robots". This dataset defines deformable objects as objects that lack a fixed shape in indoor environments, such as feces, plastic bags, socks, ropes, etc.

Download

The dataset can be downloaded from:

Google Drive

Installation

To use the dataset, you need to install MMSegmentation and SegNeXt. Please refer to the corresponding repository for installation instructions. The configuration files for the models mentioned in the paper can be found in the dos_config directory.

Train

Before training the models, you need to convert the raw annotation files in train_annotations/ and val_annotations/ to COCO or VOC format. This project uses COCO format data for model training. You can use the following scripts for conversion:

COCO format:

python labelme2coco.py ${ANNOTATION_FILENAME} ${OUTPUT_FILENAME} --labels ${CATEGORY_FILE}

VOC format:

python labelme2voc.py ${ANNOTATION_FILENAME} ${OUTPUT_FILENAME} --labels ${CATEGORY_FILE}

The script cocotomask.py converts the COCO dataset to mask images with pixel values in the range of [0,num_class-1]:

python cocotomask.py ${COCO_DATASET_FILENAME}

Statistics

The script dataset_analysis1.py counts the number of images containing a specific number of instances:

python dataset_analysis1.py ${INPUT_ANNOTATION_DIRECTORY}

The script dataset_analysis2.py provides statistics on image resolution, number of instances, number of instances for each category, and annotation area for each instance:

python dataset_analysis2.py ${COCO_ANNOTATION_JSON_FILENAME}

Inference

mscan_l.pth is the pre-trained model for SegNeXt, and best_mIoU_iter_160000.pth is the model trained on DOS Dataset using SegNeXt. They can be downloaded from:

Google Drive

You can use the example scripts video_demo.py and image_demo.py in the demo directory to perform segmentation on videos or images using SegNeXt. We also provide a sample video https://github.com/zehantan6970/DOS_Dataset/blob/main/contact_4.mp4.

Notes

For the first submission, only submit the files that need to be added to the MMSegmentation and SegNeXt.