Semantic Segmentation with MetaPrompts

Getting Started

Follow the guide in mmseg to prepare ADE20k and CityScapes datasets.

Results and Fine-tuned Models

ADE20K

Model	Config	Head	Crop Size	Lr Schd	mIoU	mIoU (ms+flip)	Fine-tuned Model
MetaPromptsSeg	config	Upernet	512x512	80K	55.83	56.81	Google drive

CityScapes w/o Mapillary pretraining

Model	Config	Head	Crop Size	Lr Schd	mIoU	mIoU (ms+flip)	Fine-tuned Model
MetaPromptsSeg	config	Upernet	1024x1024	80K	84.38	85.77	Google drive

CityScapes w/ Mapillary pretraining

Model	Config	Head	Crop Size	Lr Schd	mIoU	mIoU (ms+flip)	Fine-tuned Model
MetaPromptsSeg	config	Upernet	1024x1024	80K	85.98	87.26	Google drive

Training on ADE20K

bash dist_train.sh configs/ade.py <NUM_GPUS> --work-dir <WORK_DIR>

We use 8 GPUs by default.

Training on CityScapes w/o Mapillary pretraining

bash dist_train.sh configs/cityscapes.py <NUM_GPUS> --work-dir <WORK_DIR>

Training on CityScapes w/ Mapillary pretraining

Download the pretraining checkpoint

bash dist_train.sh configs/cityscapes_extra.py <NUM_GPUS> --work-dir <WORK_DIR> --load-from <CHECKPOINT_PATH>

Evaluation

Command format:

bash dist_test.sh configs/<>.py <CHECKPOINT_PATH> <NUM_GPUS> --eval mIoU

To evaluate a model with multi-scale and flip, run

bash dist_test.sh configs/<>_ms.py <CHECKPOINT_PATH> <NUM_GPUS> --eval mIoU