Fisheye-InternImage

The following codes are the solutions (3st place, private score: 0.66886) for the dacon competition.

If you would like to know more about the competition, please refer to the following link:

2023 Samsung AI Challenge : Camera-Invariant Domain Adaptation

• 주최 : Samsung Advanced Institute of Technology
• 주관 : DACON

Usage

Move to the "segmentation" directory and follow the README.md

Directory Structure

/workspace
├── configs
│   ├── _base_
│   │   ├── datasets
│   │   │   ├── samsung_fisheye.py
│   ├── samsung
│   │   ├── exp_01.py
│   │   ├── exp_02.py
│   │   ├── exp_03.py
│   │   ├── exp_04.py
│   │   ├── exp_05.py
├── data
│   ├── preprocess_data (will be made after running fisheye_transform.py)
│   │   ├── train_fisheye_gt
│   │   ├── train_fisheye_image
│   │   ├── val_fisheye_gt
│   │   ├── val_fisheye_image
│   ├── test_image
│   ├── train_source_gt
│   ├── train_source_image
│   ├── train_target_image
│   ├── val_source_gt
│   ├── val_source_image
│   ├── sample_submission.csv
│   ├── test.csv
│   ├── train_source.csv
│   ├── train_target.csv
│   ├── val_source.csv
├── results
│   ├── exp_04.pkl
│   ├── exp_05.pkl
├── submission
│   ├── exp_04.csv
│   ├── exp_05.csv
├── work_dirs
│   ├── exp_04
│   │   ├── best_mIoU_iter_22000.pth
├── fisheye_transform.py
├── submit.py
├── test.py
├── train.py
      .
      .
      .

Experiments

The final submission is exp_04.

Index	Model	Private mIoU	Public mIoU	Val mIoU (%)	Iter	Fine-tuned model	Pre-trained model
exp_01	UperNet	0.64483	0.6192	67.30	20000	ckpt	ckpt
exp_02	Mask2Former	0.66456	0.62771	66.01	9000	ckpt	ckpt
exp_03	UperNet	0.65114	0.62775	68.03	14000	ckpt	ckpt
exp_04	Mask2Former	0.66886	0.63133	70.34	22000	ckpt	ckpt
exp_05	Mask2Former	0.67288	0.62905	-	40000	ckpt	ckpt