Skip to content

Latest commit

 

History

History
176 lines (135 loc) · 6.58 KB

README.md

File metadata and controls

176 lines (135 loc) · 6.58 KB
Raw

Cityscapes SOTA

The implementation of Hierarchical Multi-Scale Attention based on PaddlePaddle. [Paper]

Based on the above work, we made some optimizations:

  • Use dice loss and bootstrapped cross entropy loss instead of cross entropy
  • Learn all fine data and equal amount of coarse data in each epoch
  • The evaluation is carried out by using the equal difference scale series instead of the equal ratio scale series

We achieve mIoU of 87% on Cityscapes validation set.

The actual effect is as follows (for high-definition pictures, please click here).

Installation

step 1. Install PaddlePaddle

System Requirements:

  • PaddlePaddle >= 2.0.0rc1
  • Python >= 3.6+

Highly recommend you install the GPU version of PaddlePaddle, due to large overhead of segmentation models, otherwise it could be out of memory while running the models. For more detailed installation tutorials, please refer to the official website of PaddlePaddle。

step 2. Install PaddleSeg

You should use API Calling method to install PaddleSeg for flexible development.

pip install paddleseg

Data Preparation

Download following files and put into data/cityscapes directory. Then unzip these files.

mkdir -p data/cityscapes

Firstly please download 3 files from Cityscapes dataset

  • leftImg8bit_trainvaltest.zip (11GB)
  • gtFine_trainvaltest.zip (241MB)
  • leftImg8bit_trainextra.zip (44GB)

Run the following commands to do the label conversion:

pip install cityscapesscripts
python ../../tools/data/convert_cityscapes.py --cityscapes_path data/cityscapes --num_workers 8

Where 'cityscapes_path' should be adjusted according to the actual dataset path. 'num_workers' determines the number of processes started and the size can be adjusted according to the actual situation.

Then download and uncompress Autolabelled-Data from google drive

  • refinement_final_v0.zip # This file is needed for autolabelled training for recreating SOTA

Delete useless tmp directory in refinement_final directory.

rm -r tmp/

Convert autolabelled data according to PaddleSeg data format:

python tools/convert_cityscapes_autolabeling.py --dataset_root data/cityscapes/

Finally, you need to organize data following the below structure.

cityscapes
|
|--leftImg8bit
|  |--train
|  |--val
|  |--test
|
|--gtFine
|  |--train
|  |--val
|  |--test
|
|--leftImg8bit_trainextra
|  |--leftImg8bit
|     |--train_extra
|        |--augsburg
|        |--bayreuth
|        |--...
|
|--convert_autolabelled
|  |--augsburg
|  |--bayreuth
|  |--...

Evaluation

Download Trained Model

mkdir -p saved_model && cd saved_model
wget https://bj.bcebos.com/paddleseg/dygraph/cityscapes/mscale_ocr_hrnetw48_cityscapes_autolabel_mapillary/model.pdparams
cd ..

Evaluation on Cityscapes

Model Backbone mIoU mIoU (flip) mIoU (5 scales + flip)
MscaleOCRNet HRNet_w48 86.89% 86.99% 87.00% 
CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python -u -m paddle.distributed.launch val.py \
--config configs/mscale_ocr_cityscapes_autolabel_mapillary.yml --num_workers 3 --model_path saved_model/model.pdparams

The reported mIoU should be 86.89. This evaluates with scales of 0.5, 1.0 and 2.0. This requires 14.2GB of GPU memory.

CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python -u -m paddle.distributed.launch val.py \
--config configs/mscale_ocr_cityscapes_autolabel_mapillary.yml --num_workers 3 --model_path saved_model/model.pdparams \
--aug_eval --flip_horizontal

The reported mIoU should be 86.99. This evaluates with scales of 0.5, 1.0, 2.0 and flip horizontal. This requires 14.2GB of GPU memory.

CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python -u -m paddle.distributed.launch val.py \
--config configs/mscale_ocr_cityscapes_autolabel_mapillary_ms_val.yml --num_workers 3 --model_path saved_model/model.pdparams \
--aug_eval --flip_horizontal

The reported mIoU should be 87.00. This evaluates with scales of 0.5, 1.0, 1.5, 2.0, 2.5 and flip horizontal. This requires 21.2GB of GPU memory.

Training

Download Pretrained Weights

mkdir -p pretrain && cd pretrain
wget https://bj.bcebos.com/paddleseg/dygraph/cityscapes/ocrnet_hrnetw48_mapillary/pretrained.pdparams
cd ..

Pretrained weights were obtained by pretraining on the Mapillary dataset from OCRNet (backbone is HRNet w48).

Training on Cityscapes

CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python -u -m paddle.distributed.launch train.py \
--config configs/mscale_ocr_cityscapes_autolabel_mapillary.yml --use_vdl \
--save_dir saved_model/mscale_ocr_cityscapes_autolabel_mapillary --save_interval 2000 --num_workers 5 --do_eval

Note that this requires 32GB of GPU memory. You can remove argument --do_eval to turn off evaluation during training, thus it only requires 25GB of GPU memory. If you run out of memory, try to lower the crop size.

Deploy

Run the following command to export the inference model.

python export.py \
    --config configs/mscale_ocr_cityscapes_autolabel_mapillary_ms_val.yml \
    --save_dir ./output \
    --input_shape 1 3 2048 1024

We can use the following deployment methods to deploy the inference model.

Platform Library Tutorial
Python Paddle prediction library e.g.
C++ Paddle prediction library e.g.
Mobile PaddleLite e.g.
Front-end PaddleJS e.g.

Other deployment documents:

Thanks for their contributions!