DGMN2 for Semantic Segmentation

This folder contains the implementation of DGMN2 for semantic segmentation.

Here, we take MMSegmentation as an example, applying DGMN2-Tiny to SETR-Naive, SETR-PUP head and SETR-MLA head.

Results

Method	Backbone	Iters	mIoU	mIoU (ms + flip)	Config	Download
Semantic FPN	DGMN2-Tiny	40K	78.09	79.40	config	model
Semantic FPN	DGMN2-Small	40K	80.65	81.58	config	model
Semantic FPN	DGMN2-Medium	40K	80.60	81.79	config	model
Semantic FPN	DGMN2-Large	40K	81.75	82.64	config	model
SETR-Naive	DGMN2-Tiny	40K	77.23	78.23	config	model
SETR-Naive	DGMN2-Small	40K	80.31	81.04	config	model
SETR-Naive	DGMN2-Medium	40K	80.83	81.39	config	model
SETR-Naive	DGMN2-Large	40K	81.80	82.61	config	model
SETR-PUP	DGMN2-Tiny	40K	78.25	79.26	config	model
SETR-PUP	DGMN2-Small	40K	79.78	80.73	config	model
SETR-PUP	DGMN2-Medium	40K	80.97	81.80	config	model
SETR-PUP	DGMN2-Large	40K	81.58	82.27	config	model
SETR-MLA	DGMN2-Tiny	40K	78.25	79.32	config	model
SETR-MLA	DGMN2-Small	40K	80.79	81.62	config	model
SETR-MLA	DGMN2-Medium	40K	81.09	82.00	config	model
SETR-MLA	DGMN2-Large	40K	81.55	81.98	config	model

Clone the repository locally:

git clone https://github.com/fudan-zvg/DGMN2

a. Install MMSegmentation following the official instructions. Here we use MMSegmentation 0.16.0.

b. Install PyTorch Image Models. Here we use PyTorch Image Models 0.4.5.

pip install timm==0.4.5

c. Build the extension.

cd dcn
python setup.py build_ext --inplace

First, prepare Cityscapes dataset according to the guidelines in MMSegmentation.

Then, download the weights pretrained on ImageNet, and put them in a folder pretrained/.

To train DGMN2-Tiny + SETR-PUP head on Cityscapes training set on a single node with 4 GPUs for 40K iterations run:

dist_train.sh configs/setr_pup_dgmn2_tiny_4x2_769x769_40k_cityscapes.py 4

To evaluate DGMN2-Tiny + SETR-PUP head on Cityscapes validation set on a single node with 4 GPUs run:

dist_test.sh configs/setr_pup_dgmn2_tiny_4x2_769x769_40k_cityscapes.py /path/to/checkpoint_file 4 --eval mIoU