Competition URL: https://tbrain.trendmicro.com.tw/Competitions/Details/22
Private Leaderboard: 2 / 307 (Top 1%)
In this competition, we refer to the popular model "Deeplab-v3-plus" on the segmentation task. The model architecture is shown below:
In the inference phase, we use the Test-Time Augmentation(TTA) to make our model more robust. The procedure of TTA is demonstrated below:
Here, we take a few figures from the prediction results as examples.
- Clone this repo to your local
git clone https://github.com/come880412/STAS_segmentation
cd STAS_segmentation-
System: Ubuntu20.04
-
Pytorch version: Pytorch 1.7 or higher
-
Python version: Python 3.7
-
Testing:
CPU: AMR Ryzen 7 4800H with Radeon Graphics RAM: 32GB
GPU: NVIDIA GeForce RTX 1660Ti 6GB -
Training (TWCC):
CPU: Intel(R) Xeon(R) Gold 6154 CPU @ 3.00GHz
RAM: 180 GB
GPU: Tesla V100 32 GB * 4
Please read the "requirement.txt" for the details.
- You should prepare the dataset from here, and put the dataset on the folder
../dataset. After doing so, please use the following command to do data preprocessing.
python3 preprocessing.py - Note: please modify the dataset path on the script
preprocessing.py.
- Please download the pretrained models from here, and put the models on the folder
./checkpoint.
- After downloading the pretrained models and preparing the datasets, you could use the following command to test the best results on the public/private leaderboard.
python3 test.py --root path/to/dataset --load ./checkpoint/deeplab_1280_900/model.pth --threshold 0.35- The result will be saved on the folder
./publicFig_deeplabautomatically.
- You should download the COCO pretrained models on [1]. And put the model on the folder
./pretrained. After that, please use the following script to train the best model used in this competition.
bash train.sh- Furthermore, the following figures are our model's learning curve:
- If you have any questions, please feel free to email me! come880412@gmail.com