The base code for this experiment was taken from the official open-mmlab repos
This repo contains the code for comparing the performance among a PSPHead (PPM), an ASPPHead (ASPP), and a Depth-wiseSeparableASPPHead (Depth-wise Separable ASPP) on the semantic segmentation task. For easy comparison, I use a PSPNet Encoder with these different decoder heads.
- Create a conda virtual environment and activate it.
conda create -n mm
conda activate mm
- Install PyTorch and torchvision:
conda install pytorch=1.6.0 torchvision cudatoolkit=10.2 -c pytorch
- Install mmcv requirements for Linux:
cd mmcv
pip install -e .
- Install MMSegmentation requrements
cd mmsegmentation
pip install -e .
The data could be found here after registration.
mmsegmentation
├── mmseg
├── tools
├── configs
├── data
│ ├── cityscapes
│ │ ├── leftImg8bit
│ │ │ ├── train
│ │ │ ├── val
│ │ ├── gtFine
│ │ │ ├── train
│ │ │ ├── val
- Install cityscapesscripts
pip install cityscapesscripts
- Download the zip files
# gtFine
wget --keep-session-cookies --save-cookies=cookies.txt --post-data 'username=myusername&password=mypassword&submit=Login' https://www.cityscapes-dataset.com/login/
wget --load-cookies cookies.txt --content-disposition https://www.cityscapes-dataset.com/file-handling/?packageID=1
# leftImgBit
wget --keep-session-cookies --save-cookies=cookies.txt --post-data 'username=myusername&password=mypassword&submit=Login' https://www.cityscapes-dataset.com/login/
wget --load-cookies cookies.txt --content-disposition https://www.cityscapes-dataset.com/file-handling/?packageID=3
- Unzip files into the
mmsegmentation/data/cityscapes/:
python unzip.py
- Prepare the dataset:
cd mmsegmentation/
# --nproc means 8 process for conversion, which could be omitted as well.
python tools/convert_datasets/cityscapes.py data/cityscapes --nproc 8
cd mmsegmentation/
CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 PORT=29501 ./tools/dist_train.sh configs/pspnet/pspnet_r50-d8_test_cityscapes.py 8
cd mmsegmentation/
CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 PORT=29501 ./tools/dist_train.sh configs/pspnet/pspnet_r50-d8_test_cityscapes_aspp.py 8
cd mmsegmentation/
CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 PORT=29501 ./tools/dist_train.sh configs/pspnet/pspnet_r50-d8_test_cityscapes_d_aspp.py 8
-
I conduct experiments on the Cityscapes Fine Annotations dataset.
-
For faster training and testing, I passed cropped images of size 300x300 to the models.
-
ResNet-50 is used as a backbone network.
-
The models are trained for
10000 Epochson 2975 images, with periodical testing every1000 Epochson 500 images. -
Each Model takes around
5 hoursduring training.
| Class | IoU (PPM) | IoU (ASPP) | IoU (Depth-wise ASPP) | Acc (PPM) | Acc (ASPP) | Acc (Depth-wise ASPP) |
|---|---|---|---|---|---|---|
| road | 96.45 | 97.12 | 96.68 | 97.57 | 98.25 | 97.74 |
| sidewalk | 78.2 | 80.15 | 78.83 | 90.34 | 90.08 | 90.41 |
| building | 89.29 | 89.2 | 89.0 | 94.47 | 93.02 | 93.81 |
| wall | 33.16 | 37.54 | 38.35 | 41.4 | 48.04 | 51.38 |
| fence | 52.93 | 53.63 | 53.8 | 67.58 | 69.93 | 70.2 |
| pole | 57.13 | 57.93 | 57.6 | 68.19 | 69.87 | 67.51 |
| traffic light | 63.92 | 61.78 | 61.45 | 77.72 | 77.37 | 79.55 |
| traffic sign | 72.36 | 70.49 | 71.29 | 82.57 | 83.31 | 81.67 |
| vegetation | 91.14 | 90.75 | 91.17 | 96.29 | 96.66 | 96.05 |
| terrain | 60.41 | 59.19 | 55.26 | 72.82 | 71.83 | 72.47 |
| sky | 92.5 | 91.0 | 91.49 | 97.34 | 97.81 | 96.2 |
| person | 77.19 | 77.07 | 77.9 | 88.5 | 89.38 | 89.23 |
| rider | 51.03 | 50.96 | 51.28 | 65.08 | 66.43 | 65.38 |
| car | 92.3 | 92.68 | 92.37 | 97.2 | 97.22 | 97.09 |
| truck | 46.44 | 42.97 | 38.32 | 56.71 | 58.99 | 44.6 |
| bus | 61.7 | 65.2 | 56.5 | 71.36 | 77.89 | 76.43 |
| train | 37.39 | 39.27 | 25.71 | 76.65 | 78.06 | 74.93 |
| motorcycle | 55.29 | 56.21 | 46.03 | 68.57 | 75.01 | 56.09 |
| bicycle | 72.83 | 72.62 | 71.63 | 88.13 | 86.35 | 87.83 |
| Scope | mIoU (PPM) | mIoU (ASPP) | mIoU (Depth-wise ASPP) | mAcc (PPM) | mAcc (ASPP) | mAcc (Depth-wise ASPP) |
|---|---|---|---|---|---|---|
| global | 67.46 | 67.67 | 65.51 | 78.87 | 80.29 | 78.35 |