Could you provide script to train on BDD100K for segmentation task?

[John1231983]

Great job. Could you also provide the dataset loader and script to train on your dataset (BDD100K) for segmentation? Thanks

[fyu]

The command is almost the same with Cityscapes training. Because images in BDD100K are 720p, I use 640 for crop size. So the command is like

python3 segment.py train -d <data_folder> -c <category_number> -s 640 \
    --arch drn_d_22 --batch-size 32 --epochs 250 --lr 0.01 --momentum 0.9 \
    --step 100

The models trained on BDD100K will be added to this repo.

[isht8]

what average IoU does your model acheive when trained(for semantic segmentation) on the 19 classes of BDD100k?

[yyfyan]

I had trained the semantic segmentation model on BDD100K,now I want to test the model on the test set. How to test it? @fyu

[AmeetR]

@yyfyan Did you write your own script for cleaning up the bdd datasets? I think it would be pretty simple to effectively copy the cityscapes preprocessing but using the BDD structure, but want to make sure.

[abhigoku10]

@John1231983 @yyfyan @fyu Thanks for open sourcing the source code , its a great work . I have few queries
Q1 when i used the pre trained model on BDD100k and on custom dataset i get the below results
Q2 should we use any different kind of preprocessing technique on bdd100k and custom dataset
to obtain results like cityscapes
Q3 any idea to increase the mask quality for poles and get the trunk part of tree more clearly
Q4 y does sky getting classified and segmented as buildings or trees

THanks in advance

[parthkvv]

@abhigoku10 Did you train a model on a custom dataset? I am trying to do the same and need some guidance. Thanks in advance.