Semantic Segmentation with HRNet + OCR on MORAI dataset

This repository is based on HRNet + OCR and mmsegmentation. All configurations and codes were revised for MORAI dataset.

Results and Models

HRNet + OCR

Class accuracy is IoU(Intersection over Union).

Dataset	Iter	vehicle	bus	truck	freespace	roadMark	whiteLane	yellowLane	trafficSign	fence	background	config	log	model
Real	160k	93.29	89.92	87.72	94.94	61.22	87.81	79.86	66.3	91.08	97.85	config	log	model
Daegu	160k	93.16	71.36	87.35	98.03	82.69	83.64	84.3	71.6	83.87	99.58	config	log	model
Sejong BRT 1	160k	55.83	75.51	53.52	98.67	66.46	82.2	63.77	72.14	X	99.83	config	log	model
Sangam Edge	160k	84.57	0.0	0.0	97.77	82.93	79.45	84.59	91.56	92.47	99.03	config	log	model
Sejong BRT 1 Edge	160k	53.91	0.0	0.0	98.4	X	92.01	35.98	0.0	X	99.56	config	log	model

Some of the classes not included in data or in small amounts.

Real:

Daegu:

Sejong BRT 1:

Sangam Edge:

Sejong BRT 1 Edge:

Mixed Models(10% real + 90% synthetic)

Dataset	Iter	vehicle	bus	truck	freespace	roadMark	whiteLane	yellowLane	trafficSign	fence	background	config	log	model
Daegu	160k	89.72	67.8	64.09	95.72	40.56	83.04	75.65	41.92	63.48	97.16	config	log	model
Sejong BRT 1	160k	86.25	77.94	62.12	95.82	49.5	82.73	74.57	37.21	67.79	97.68	config	log	model
Sangam Edge	160k	89.55	68.71	65.34	95.9	49.17	84.86	73.56	38.53	80.18	97.96	config	log	model
Sejong BRT 1 Edge	160k	84.95	68.66	20.02	94.88	0.94	82.81	72.19	0.0	49.82	94.42	config	log	model

Some of the classes not included in data or in small amounts.

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

NOTE: Change mmseg/datasets/custom.py if original image format is in .jpg

…
def __init__(self,
                 pipeline,
                 img_dir,
                 # img_suffix='_leftImg8bit.png',
                 img_suffix='.png', -> Modify into ‘.jpg’
…

# single-gpu testing
python tools/test.py {CONFIG_FILE} {MODEL_FILE} --eval mIoU \
(--show-dir {LOCATION})

# multi-gpu testing
(CUDA_VISIBLE_DEVICES={GPU_NUM}) \
tools/dist_test.sh {CONFIG_FILE} {MODEL_FILE} {TOTAL_NUM_OF_GPU} --eval mIoU \
(--show-dir {LOCATION})

--show-dir saves pictures of result. You can use --show in GUI environment.

CUDA_VISIBLE_DEVICES can define which GPUs will be used. If not defined, they are used sequentially.

Example:

python tools/test.py \
configs/ocrnet/ocrnet_hr48_512x1024_160k_morai_daegu.py \
checkpoints/ocrnet_hr48_512x1024_160k_morai_daegu.pth \
--eval mIoU --show-dir result.mIoU.daegu/

CUDA_VISIBLE_DEVICES=0,1,3 tools/dist_test.sh \
configs/ocrnet/ocrnet_hr48_512x1024_160k_morai_daegu.py \
checkpoints/ocrnet_hr48_512x1024_160k_morai_daegu.pth 3 \
--eval mIoU --show-dir result.mIoU.daegu/

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/ocrnet/ocrnet_hr48_512x1024_160k_morai_daegu.py

CUDA_VISIBLE_DEVICES=0,1,3 tools/dist_train.sh \
configs/ocrnet/ocrnet_hr48_512x1024_160k_morai_daegu.py 3