2D Object Detection with Swin Transformer on MORAI dataset

This repository is based on Swin-Transformer-Object-Detection and mmdetection. All configurations and codes were revised for MORAI dataset.

Results and Models

Swin-L + FPN + Cascade R-CNN

Dataset	Epoch	box AP(vehicle)	config	log	model
Real	36	85.8	config	log	[model]
Daegu	36	68.3	config	log	[model]
Sejong BRT 1	36	70.5	config	log	[model]
Sangam Edge	36	71.1	config	log	[model]
Sejong BRT 1 Edge	36	69.6	config	log	[model]

Real:

Daegu:

Sejong BRT 1:

Sangam Edge:

Sejong BRT 1 Edge:

Mixed Models(10% real + 90% synthetic)

Dataset	Epoch	Real test-set box AP(vehicle)	config	log	model
Daegu	36	73.9	config	log	[model]
Sejong BRT 1	36	71.3	config	log	[model]
Sangam Edge	36	70.3	config	log	[model]
Sejong BRT 1 Edge	36	64.8	config	log	[model]

Daegu:

Sejong BRT 1:

Sangam Edge:

Sejong BRT 1 Edge:

Usage

Installation

Please refer to install.md for installation, dataset preparation and making configuration file.

Testing, Demo

# single-gpu testing
python tools/test.py {CONFIG_FILE} {MODEL_FILE} --eval bbox \
(--show-dir {LOCATION}) \
(--options "classwise=True")

# multi-gpu testing
(CUDA_VISIBLE_DEVICES={GPU_NUM}) \
tools/dist_test.sh {CONFIG_FILE} {MODEL_FILE} {TOTAL_NUM_OF_GPU} --eval bbox \
(--show-dir {LOCATION}) \
(--options “classwise=True”)

--show-dir saves pictures of result, --options "classwise=True" shows average precision of all classes. You can use --show in GUI environment.

Example:

python tools/test.py \
configs/swin/cascade_mask_rcnn_swin_base_patch4_window7_mstrain_480-800_giou_4conv1f_adamw_3x_morai_daegu.py \
checkpoints/cascade_mask_rcnn_swin_base_patch4_window7_mstrain_480-800_giou_4conv1f_adamw_3x_morai_daegu.pth \
--eval bbox --show-dir result.bbox.daegu/ --options “classwise=True”

CUDA_VISIBLE_DEVICES=0,1,3 tools/dist_test.sh \
configs/swin/cascade_mask_rcnn_swin_base_patch4_window7_mstrain_480-800_giou_4conv1f_adamw_3x_morai_daegu.py \
checkpoints/cascade_mask_rcnn_swin_base_patch4_window7_mstrain_480-800_giou_4conv1f_adamw_3x_morai_daegu.pth 3 \
--eval bbox --show-dir result.bbox.daegu/ --options “classwise=True”

Training

# single-gpu training
python tools/train.py {CONFIG_FILE}

# multi-gpu training
(CUDA_VISIBLE_DEVICES={GPU_NUM}) tools/dist_train.sh {CONFIG_FILE} {TOTAL_NUM_OF_GPU}

Example:

python tools/train.py configs/swin/cascade_mask_rcnn_swin_base_patch4_window7_mstrain_480-800_giou_4conv1f_adamw_3x_morai_daegu.py

CUDA_VISIBLE_DEVICES=0,1,3 tools/dist_train.sh \
configs/swin/cascade_mask_rcnn_swin_base_patch4_window7_mstrain_480-800_giou_4conv1f_adamw_3x_morai_daegu.py 3