shakibyzn / Cityscapes_MMSegmentation

Cityscapes_MMSegmentation Project

This project involves preparing the Cityscapes dataset, compressing it with different JPEG quality levels, and evaluating the mIoU using the DeepLabV3 with ResNet50 pre-trained checkpoint from MMSegmentation.

Steps to Reproduce

Step 1: Download Cityscapes Dataset

1. Sign up and create an account on the Cityscapes dataset website.
2. Download the following files:
   • gtFine_trainvaltest.zip (241MB) -> Ground truth
   • leftImg8bit_trainvaltest.zip (11GB) -> Training RGB Images

Step 2: Compress Cityscapes Dataset with JPEG

Use ImageMagick to compress the dataset with two different quality levels:

• Q=0.25
• Q=0.75

Example command:

$ bash compress.sh 25%

Step 3: Training and Inference

Follow the steps from the notebook file.

Evaluation Results

Here are the evaluation results based on the specified method:

Method	aAcc	mIoU	mAcc
Inference-25%	38.07	5.81	8.60
Fine-tuning-25%	78.49	19.49	25.41
Inference-75%	34.77	5.00	9.58
Fine-tuning-75%	70.20	14.47	20.28