Deeplabv3+ BDD100k/drivable_area implementation

Introduction

This repository is from https://github.com/jfzhang95/pytorch-deeplab-xception. For BDD100k/drivable_area semantic segmentation, I added

1. bdd100k drivable area dataloader, and training/val/test scripts.
2. prediction visualization for both color (visual result) and id (greyscale png file for submission).
3. replaced Batch Normalization with Group Noramlization.

For more detail, please visit the repository above.

How to train

bash go_train.sh

If you want to specify cropsize, please check the code 'dataloaders/datasets/bdd100k.py'. Default is non-cropping input.

How to evaluate

bash go_eval.sh

How to inference

For visual result,

bash go_inference.sh

The code results out only 100 example, but you can easily change the code.

For submission,

bash go_submit.sh

Experiment & result

0. Architecture

I used Resnet101 as backbone. Used ImageNet pretrianed model, and finetuned. I replaced all Batch Normalization layers with Group Normalization layer which groups 16 channels.

1. Hyperparameter & environment

I did not crop, but fed original image as input. Replaced every BN with Group Nomralization. Compared to BN, I could use more smaller batchsize, feeding higher resolution and non-cropped image.

Learning rate: 0.01 / 16 * batchsize

weight_decay: 0.0005

momentum: 0.9

max epoch: 30

Data augmentation: random gaussian blur for training, normalize with BDD100k mean and std (not that of ImageNet) for both training and validation.

Single 12GB GPU

Backbone	train/eval os	mIoU in val	mIoU in test
ResNet101	16/16	85.50%	85.33%

WAD2018	Score	Difference
1st	86.18	-0.85
2nd	86.04	-0.71
Mine	85.33	+0.0
3rd	84.01	+1.32