/SAM_Objects365

Use the SAM to get the mask on Objects-365.

Primary LanguagePythonApache License 2.0Apache-2.0

SAM-Objects365

TODO

  • Release the download links.
  • Release the toturial.

Introduction

The goal of this project is to use SAM to annotate masks on the basis of carefully annotated target detection datasets Objects365, thereby facilitating large-scale pre-training of general instance segmentation. The specific process is shown in the figure below

Visualizations

We provide some annotations of our SAM-Objects365. More visualizations please refer to this link.

How to use SAM-Objects365 ?

Download

Firstly, you should download the Objects365 v2 by running following:

python download_dataset.py \
    --dataset-name objects365v2 \
    --save-dir ${SAVING PATH} \
    --unzip \
    --delete  # Optional, delete the download zip file

Next you need download following mask annotations:

BaiduNetDisk: https://pan.baidu.com/s/1KLncoQ9u7-ABrgyBKCwxGg (1234)

OneDrive: https://1drv.ms/f/s!AqU-46ABHrbCgc8bS9HfrrYjthzf0Q?e=6CHGjs

This dataset folder sholud be like:

objects365_v2
├── annotations
│   ├── sam_obj365_train_1742k.json
│   ├── sam_obj365_train_75k.json
│   ├── sam_obj365_val_5k.json
│   ├── zhiyuan_objv2_train.json
│   └── zhiyuan_objv2_val.json
├── sam_mask_json
│   ├── sam_obj365_train_1742k
│   ├── sam_obj365_train_75k
├── train
└── val

Acknowledgement

We would like to express our heartfelt thanks for the following projects: