Open World Object Detection

This project contains Pytorch-based object detection for open world.
We purpose to identify and characterize unknown instances to assist relieveing confusion within an incremental object detection framework.

Tutorial

We provide GUI on the web page for convenience of using our methods.
run app.py and access the guided web page.
If you want to match your settings with our development environment, refer the requirements.txt.

1. Real-time Open world object detection

• If you want to use real-time Open world object detection by webcam, press Start Webcam.

2. Make YAML with new label

• If you want to create new label to the unknown label, upload unkown label.
• and fill in the new label name, press Download YAML.
• The label of new name must be ENGLISH only.

3. Train new model

• If you want to train new model, press TRAIN 시작.

• Press Unknown Label Upload and Download YAML first if you train new model with unknown image.

• Once train is terminated, required to restart app.

• Depending on the user's environment some settings should be modifed appropriately:

  • The unknown config and config in flask_detect.py
  • The similarconf in app.py

Project structure

/openset_object_detection
├── data
├── templates
├── utils
├── app.py
├── benchmarks.py
├── detect.py
├── export.py
├── hubconf.py
├── train.py
├── val.py
├── requirements.txt
├── setup.cfg
└── README.md

Face Domain Adaption

This project contains Pytorch-based domain adaptation for face recognition. We aim to solve two typical real-world face recognition problems: masked face recognition, and corrupted face recognition. To solve this challenging task, we use various domain adaptation methods (DANN, CDAN, BSP, etc).

Tutorial

We provide a basic command to train/test our method on the masked face recognition dataset (MFR1).

• To see each method's detailed configuration, please see configs/.
• To see each dataset's specification, please see docs/Evaluation.txt.

1. train

# checkpoint is saved into pretrained/DABase_IResNet/Unmasked+Synthetic_Masked_backbone.ckpt
python3 main.py fit --config=configs/corrupt/source_only.yaml -g '0,' --seed_everything 2021

2. test

# modify model-init_args-backbone_init-pretrained_path to the trained checkpoint (configs/source_only.yamsl)
python3 main.py test --config=configs/corrupt/source_only.yaml -g '0,' --seed_everything 2021

Experiment setup

category	value
model	iResNet50
epochs	25
dropout	0.1
image size	112, 112
learning rate	0.001
batch size	32
optimizer	SGD (momentum=0.9, nesterov=True)

Project structure

# project structure under `/source_free_domdain_adaptation`
/source_free_domain_adaptation
├── LICENSE
├── README.md
├── config.yaml
├── configs
├── data
├── docs
├── main.py
├── save_backbone.py
├── src
└── tmuxp

Dataset structure

# dataset structure under `/data` folder.
├── jpg
│   ├── MFR1
|   |   ├── Masked
|   |   ├── Synthetic
|   |   ├── Unmasked
|   |   ├── Unmasked+Synthetic
|   |   ├── dataset_info.txt
|   |   ├── pairs
|   |   ├── split.py
|   |   ├── split_class_dict.json
|   |   └── split_class_dict_info.txt
├── test
│   ├── MFR1
│   │   ├── Masked
│   │   ├── Unmasked
│   │   ├── m_m.bin
│   │   ├── m_m_split.bin
│   │   ├── u_m.bin
│   │   ├── u_m_split.bin
│   │   ├── u_u.bin
│   │   └── u_u_split.bin

Requirements

we test this project on the following environments:

• ubuntu linux == 18.04
• python >= 3.7.5
• cuda >= 11.3
• pytorch >= 1.10.0